WWW.JAQM.

RO

JOURNAL
OF
APPLIED
QUANTITATIVE
METHODS

Quantitative Methods in Audit and Control Vol. 4

No. 4
Winter
2009
ISSN 1842–4562
 Editorial Board

JAQM Editorial Board

Editors
Ion Ivan, University of Economics, Romania
Claudiu Herteliu, University of Economics, Romania
Gheorghe Nosca, Association for Development through Science and Education, Romania

Editorial Team
Cristian Amancei, University of Economics, Romania
Sara Bocaneanu, University of Economics, Romania
Catalin Boja, University of Economics, Romania
Irina Maria Dragan, University of Economics, Romania
Eugen Dumitrascu, Craiova University, Romania
Matthew Elbeck, Troy University, Dothan, USA
Nicu Enescu, Craiova University, Romania
Bogdan Vasile Ileanu, University of Economics, Romania
Miruna Mazurencu Marinescu, University of Economics, Romania
Daniel Traian Pele, University of Economics, Romania
Ciprian Costin Popescu, University of Economics, Romania
Marius Popa, University of Economics, Romania
Mihai Sacala, University of Economics, Romania
Cristian Toma, University of Economics, Romania
Erika Tusa, University of Economics, Romania
Adrian Visoiu, University of Economics, Romania

Manuscript Editor
Lucian Naie, SDL Tridion

I
 Advisory Board

JAQM Advisory Board

Luigi D’Ambra, University “Federico II” of Naples, Italy

Ioan Andone, Al. Ioan Cuza University, Romania
Kim Viborg Andersen, Copenhagen Business School, Denmark
Tudorel Andrei, University of Economics, Romania
Gabriel Badescu, Babes-Bolyai University, Romania
Catalin Balescu, National University of Arts, Romania
Avner Ben-Yair, Sami Shamoon Academic College of Engineering, Israel
Constanta Bodea, University of Economics, Romania
Ion Bolun, Academy of Economic Studies of Moldova
Recep Boztemur, Middle East Technical University Ankara, Turkey
Constantin Bratianu, University of Economics, Romania
Irinel Burloiu, Intel Romania
Ilie Costas, Academy of Economic Studies of Moldova
Valentin Cristea, University Politehnica of Bucharest, Romania
Marian-Pompiliu Cristescu, Lucian Blaga University, Romania
Victor Croitoru, University Politehnica of Bucharest, Romania
Cristian Pop Eleches, Columbia University, USA
Michele Gallo, University of Naples L'Orientale, Italy
Bogdan Ghilic Micu, University of Economics, Romania
Anatol Godonoaga, Academy of Economic Studies of Moldova
Alexandru Isaic-Maniu, University of Economics, Romania
Ion Ivan, University of Economics, Romania
Radu Macovei, University of Medicine Carol Davila, Romania
Dumitru Marin, University of Economics, Romania
Dumitru Matis, Babes-Bolyai University, Romania
Adrian Mihalache, University Politehnica of Bucharest, Romania
Constantin Mitrut, University of Economics, Romania
Mihaela Muntean, Western University Timisoara, Romania
Ioan Neacsu, University of Bucharest, Romania
Peter Nijkamp, Free University De Boelelaan, The Nederlands
Stefan Nitchi, Babes-Bolyai University, Romania
Gheorghe Nosca, Association for Development through Science and Education, Romania
Dumitru Oprea, Al. Ioan Cuza University, Romania
Adriean Parlog, National Defense University, Bucharest, Romania
Victor Valeriu Patriciu, Military Technical Academy, Romania
Perran Penrose, Independent, Connected with Harvard University, USA and London University, UK
Dan Petrovici, Kent University, UK
Victor Ploae, Ovidius University, Romania
Gabriel Popescu, University of Economics, Romania
Mihai Roman, University of Economics, Romania
Ion Gh. Rosca, University of Economics, Romania
Gheorghe Sabau, University of Economics, Romania
Radu Serban, University of Economics, Romania
Satish Chand Sharma, Janta Vedic College, Baraut, India
Ion Smeureanu, University of Economics, Romania
Ilie Tamas, University of Economics, Romania
Nicolae Tapus, University Politehnica of Bucharest, Romania
Timothy Kheng Guan Teo, National Institute of Education, Singapore
Daniel Teodorescu, Emory University, USA
Dumitru Todoroi, Academy of Economic Studies of Moldova
Nicolae Tomai, Babes-Bolyai University, Romania
Viorel Gh. Voda, Mathematics Institute of Romanian Academy, Romania
Victor Voicu, University of Medicine Carol Davila, Romania
Vergil Voineagu, University of Economics, Romania

II
 Contents

Page
Editors’ Note – JAQM 2008 Awards Announcement 406

Quantitative Methods in Audit and Control

Daniela Ioana SANDU
Multidimensional Model for the Master Budget 408
Delia BABEANU, Valerica MARES
Standards Review on Mission of Management Information Systems Audit 422
Vladimir SIMOVIC, Vojkan VASKOVIC, Dusan POZNANOVIC
A Model of Credit Bureau in Serbia – Instrument for Preserving Stability of the 429
Banking Sector in Conditions of the Global Economic Crisis

Software Analysis

Lucio T. DE PAOLIS, Marco PULIMENO, Giovanni ALOISIO

Different Simulations of a Billiards Game 440
Emil BURTESCU
Database Security - Attacks and Control Methods 449
Quantitative Methods Inquires
John P. FELDMAN, Ron GOLDWASSER, Shlomo MARK,
Jeremy SCHWARTZ, Itzhak ORION
A Mathematical Model for Tumor Volume Evaluation using Two-Dimensions 455
Chin-Lien WANG, Li-Chih WANG
A Forecasting Model with Consistent Adjustments 463
for Anticipated Future Variations

Dima ALBERG, Mark LAST, Avner BEN-YAIR

Induction of Mean Output Prediction Trees 485
from Continuous Temporal Meteorological Data

Oana STANCU (DINA), Iulia TEODORESCU (GABOR),

Catalina Liliana ANDREI, Emanuel RADU, Octavian ZARA
The Impact of Pacing Mode and the Anatomical Position 496
of Pacing Lead on the Incidence of Heart Failure

Antonio LUCADAMO, Giovanni PORTOSO

Different Approaches using the Normal and the Exponential Distribution 505
in the Evaluation of the Customer Satisfaction

Lindita ROVA, Romeo MANO

The Impact of Financial Crisis on the Quality of Life 514
Claudiu HERTELIU, Alexandru ISAIC-MANIU
Statistical Indicators for Religious Studies: Indicators of Level and Structure 525

III
 Contents

Page

Book Review

Valerica MARES
Book Review on AUDIT AND INFORMATION SYSTEMS CONTROL 532
(“AUDITUL SI CONTROLUL SISTEMELOR INFORMATIONALE”), by Pavel NASTASE
(coord.), Eden ALI, Floarea NASTASE, Victoria STANCIU, Gheorghe POPESCU,
Mirela GHEORGHE, Delia BABEANU, Dana BOLDEANU, Alexandru GAVRILA, Ed.
Economica, Bucharest, 2007

IV
 Editors’ Note – JAQM 2008 Awards

Ion IVAN
PhD, University Professor, Department of Economic Informatics
Faculty of Cybernetics, Statistics and Economic Informatics
University of Economics, Bucharest, Romania

E-mail: [email protected] ; Web Page: http://www.ionivan.ro

Claudiu HERTELIU
PhD, Assistant Professor, Department of Statistics and Econometrics,
University of Economics, Bucharest, Romania

E-mail: [email protected], Web page: http://www.hertz.ase.ro

Gheorghe NOSCA
PhD, Association for Development through Science and Education, Bucharest, Romania

E-mail: [email protected]

We are happy to announce the winners of JAQM 2008 Awards.

After deliberations, the winners are:

1st Category

For his very important contribution in promoting and developing Quantitative Methods in
Academia, as a recognized scholar, we grant to:

Régis BOURBONNAIS
from Université de Paris-Dauphine

the 2008 JAQM Medal of Honor

2nd Category

For the most valuable, Quantitative Methods related, paper published in JAQM,
we granted to:

Alexandru ISAIC-MANIU Viorel Gh. VODA

from University of Economics, and from Mathematics Institute of Romanian
Bucharest, Romania Academy, Bucharest, Romania

the 2008 JAQM Best Paper Award

406
 Editors’ Note – JAQM 2008 Awards

3rd Category

For the most promising young researcher in Quantitative Methods area, we grant to:

Marius Emanuel POPA

from University of Economics, Bucharest, Romania

the 2008 JAQM Distinction

407
 Quantitative Methods in Audit and Control

MULTIDIMENSIONAL MODEL
FOR THE MASTER BUDGET

Daniela Ioana SANDU1

PhD Candidate at the University of Economics, Bucharest, Romania
Senior BI Consultant at IBM, Milan, Italy

E-mail: [email protected], [email protected]

Abstract: In a dynamic business environment characterized by extreme competitiveness and

the need to quickly adapt to new and changing market condition, information has became an
asset. Timely and quality information are the basis for quality decisions, and only quality
decisions help survive and prosper on the market. Business intelligence applications help the
management take quality decisions. Business Performance Management applications steer the
entire organization in the same direction, enabling the organization to translate strategies into
plans, monitor execution, and provide insight to improve both financial and operational
performance. A BPM implementation often combines financial with non-financial metrics that
can identify the health of an enterprise from a variety of perspectives. BI and BPM applications
implement multidimensional models, powerful models for data analysis and simulation. The
present paper describes a multidimensional model that supports the construction of the master
budget of an enterprise with simulation facilities.

Key words: business performance management; business intelligence; planning; budgeting

1. Business Performance Management

Wayne Eckerson from The Data Warehousing Institute defines BPM as being “a
series of processes and applications designed to optimize the execution of business strategies”,
while Lee Geishecker, research director of the well-known research institution Gartner Inc
defines BPM as being a set of „ methodologies, metrics, processes and systems used to
monitor and manage an enterprise’s business performance”.
The term Business Performance Management – BPM is synonym with the terms
Corporate Performance Management - CPM , Enterprise Performance Management - EPM and
means steering the organization in the same direction by allowing the transformation of
strategies in plans, monitoring the plans execution and offering detail information on the
organization evolution for improving the enterprise’s operational and financial
performances. A BPM implementation combines financial with non-financial metrics in order
to identify the degree of health of the economic organization from a variety of perspectives.
A BPM software solution contains standard components such as:
• Planning, budgeting and forecasting: components that allows defining plans,
creating budgets and forecasts;

408
 Quantitative Methods in Audit and Control

• Financial and statutory consolidation: a component that allows realizing the

consolidated balance sheet at the level of a group of companies belonging to the
same holding;
• Scorecarding: a component that allows defining performance metrics and
dashboards;
• Reporting and analysis;
• Business Intelligence (BI).

Strategic
Planning

Forecast Budgeting

BI

Analysis
Consolidation

Reporting

Figure 1. The BPM process

2. Planning, budgeting and forecasting

A budget is the translation of strategic plans into measurable quantities that express
the expected resources required and anticipated returns over a certain period.
As defined in [8], there are several types of budgets:
• Short term (month to month, year to year) versus long term budgets (5 years);
• Fixed versus rolling budgets. A fixed budget covers a specific time frame, usually one
fiscal year. At the end of the year a new budget is prepared for the following year. A
rolling budget is a plan that is continually adapted so that the time frame remains
stable while the actual period covered by the budget changes (for example: as each
month passes, the one-year rolling budget is extended by one month so that there is
always a one-year budget in place).
• Incremental versus zero-based budgeting. Incremental budgets extrapolate from
historical figures: in determining the budget for the next period managers look at the
previous period’s budget and actuals (past figures are increased by a set percentage
or by an absolute value). In zero-based budgets each new budgeting cycle starts
from a zero base or from the ground up as though the budget were prepared for the
first time. Zero-based budgets require managers to perform a much more in-depth
analysis of each line item.

409
 Quantitative Methods in Audit and Control

The budgeting process produces the The MASTER BUDGET

master budget. The master budget brings all
the pieces together incorporating the operating
budget and the financial budget:
Operating Financial
- the operating budget consists of budget budget
budgets from each function (research
and development, production,
Development Capital budget
marketing, distribution, customer
and research
service) and provides the budgeted
income statement;
- the financial budget includes capital Design Cash budget
budget, the cash budget, the budgeted
balance sheet, the budgeted cash flow. Production
The master budget integrates Balance Sheet
operational and financial budgets and is budget
created with an iterative process during which Marketing
information flows back and forth from each Cash Flow
element of the master budget. budget
Distribution
Planning is a strategic prediction of
business performance at a summary level.
Plans are defined by senior managers who Customer
Service
help the company respond to changing market
conditions and opportunities. The process is
frequent and must be completed quickly.
Figure 2. The Master budget components
Budgeting is planning distributed to
individual areas of responsibility across the
business. Many more people are involved in the process and the work is done at a greater
work of detail. Budgeting is a slow process, often taking weeks and performed once or twice
a year.
Forecasting is a revision of the budget (sometimes in at a summarized detail) to
reflect changing market conditions, strategic plan alterations, error corrections, revised
assumptions in the original approved budget. Organizations typically re-forecast monthly; a
handful of finance personnel take part in the process.

Table 1. Planning, budgeting and forecasting (source: [8]2)

Frequency Speed Detail level Personnel involved
Planning Often Quick Summary Senior management
Budgeting Annual Slow Highly detailed All departments
Forecasting Monthly Quick Summary, light Finance
detailed

Budgeting performs 4 basic functions, each critical to a company in achieving its

strategic objectives: planning, coordinating and communicating, monitoring progress,
evaluating performance.
Planning, budgeting and forecasting are the most important management
functions. Their main purpose is to enable senior managers to see the financial implications
of various business scenarios. It is a continuous and rapid cycle that provides a near-real-

410
 Quantitative Methods in Audit and Control

time response and the most likely business plan scenario becomes the target for the
upcoming budget cycles.
Every year companies invest substantially to create comprehensive plans, an annual
budget and several forecasts, spending heavily for specialised software, staff overtime and
temporary help for data entry. Of these, budgeting it is the most difficult task. Senior
managers, accountants, financial analysts, department managers spend countless hours in
budget preparation, revision and consolidation. The overall result is as follows:
• organizations spend more time creating a budget than analyzing it;
• most times the budget bears little or no relation to the organization’s business plan;
• after the budget it is approved no one looks at it again;
• budget holders dislike very much the tedious and lengthy process of creating,
revising and submitting documents;
• budget holders usually attribute adverse variances to the finance department and
favourable variable to their own performances and managerial skills. The practice is
known as slack or padding and it occurs when managers believe they are going to
be evaluated on their performance relative to budget. To ensure that they will
achieve their budgeted figures and be rewarded, the budget revenues conservatively
or exaggerate costs or do both.
Most companies use spreadsheets as main budgeting tool. Though spreadsheets
are personal productivity tools, they have numerous shortcomings that prevent them from
adequately manage a budgeting process of any significant size or sophistication:
• spreadsheets are two-dimensional while budgeting itself is a multidimensional
process (example: budget revenue by customer, product, period, version etc);
• spreadsheets are very hard to maintain. Speed and ease in updating a budgeting
model is essential for staying abreast of business change. A simple change (adding
a cost centre, a department) can mean updating hundreds of spreadsheets and
macros.
• Spreadsheets don’t integrate well with other systems. A spreadsheet is a single-user
tool. With spreadsheets is difficult to share data with other systems (ERP’s , OLTP’s);
• Spreadsheets models are difficult to share. A spreadsheet is a single-user tool. With
spreadsheets is difficult to share data among different worksheets and workbooks.
Building a spreadsheet-based solution that consolidates input from multiple users is
tedious, time-consuming, very difficult to change and maintain.
• Spreadsheet models are hard to understand: chasing cell references around a
spreadsheet or workbook to understand one formula is a frustrating process.
Software solutions for budgeting have to provide flexibility to accurately model the
business, support multiple users, easily adapt to rapid change. Budgeting software solutions
have to address the disadvantages of spreadsheets-based systems:
• should support multidimensional budgeting;
• allow fast adaptation to changing constraints, assumptions and structures;
• have data import and export functionalities;
• be easy to use for non-programmers (business users should be able to build their
own models with IT department intervention);
• should allow calculations for simulations and what-if scenarios.
Leaders on the BPM market, all software providers offer planning, budgeting and
forecasting solution:

411
 Quantitative Methods in Audit and Control

• Hyperion Solutions (bought by Oracle in 2008) has Hyperion Planning;

• Cognos (bought by IBM in 2008) has Cognos 8 Planning;
• SAP has SAP SEM Intelligence (BPS);

3. Multidimensional models

A model is an abstraction of the real world. Models allow business analysts to give
a form, a shape to the unknown concepts and realities. The goal of a data model is to
represent in an exhaustive manner the data and information of an organization.
Legacy or OLTP (On-Line Transaction Processing) systems are the organization’s
operational systems: they are transaction oriented and used to managing day by day
activities (input or update of orders, sales, inventory, accounting records etc). These systems
use databases that implement relational models in a normalized form. A relational model is
composted by:
• Entities: are objects the organization needs to analyze.
• Relationships: describe the way entities interact with each other.
• Attributes are characteristics of the entities.
• Operations such as insert, delete and update are very fast due to the normalized
form of the database that allows minimum redundancy.

Figure 3. The relational model

Things are not as easy with data retrieval in a relational model. Retrieving the
information of a query (the best sold products on all markets during the last 3 months)
usually involves joining of more tables in order to find all the necessary data (in real
applications joins of 10 - 20 tables are very common). The more tables involved in the query,
the more data in every individual table, the more the aggregations to make (calculations like
sum, average, count etc), the longer it takes to the query to retrieve the final result. The
relational model is not fit for querying. The solution to this problem is a multidimensional
model.
The multidimensional data model enforces simplicity by giving up to the minimum
redundancy: opposite to the relational model, the multidimensional model is highly de-
normalized. De-normalization and redundancy contribute to quick retrieve time because the
information doesn’t have to be built up from a large number of tables connected by joins.
A multidimensional model is made of two types of tables:

412
 Quantitative Methods in Audit and Control

• A fact table: containing the measures, elements that are subject to analysis (sold
quantity, price ) and on which the query is build;
• Dimension tables: containing the elements on which data is to be aggregated, the
analysis axis of the data (product, client, market etc).
The dimension tables (de-normalized) are linked to the fact tables with foreign key
forming a star-schema. Sometimes the dimension tables are further normalized by
eliminating the low-cardinality attributes in separate tables: the result is a snow-flake
schema. When multiple fact tables share common dimensions we have a multi star schema.

Star Snowflake Multistar

schema schema schema

Figure 4. Multidimensional models

A star-schema forms a structure called cube. Despite its name which is a direct
expression of the limited capacity of the human brain for space representation, a cube is
made of n dimensions.
Dimensions are the analysis axis of data and they are made up of members
organized in hierarchies based on father - child relationships. A cube contains a cell for any
member combination of its dimensions.

Figure 5. Rubik’s cube: a 3 dimensional cube

One of the dimensions is the measures dimension: measures are the key
performance indicators that the business analysts want to evaluate. To determine which of
the numbers in the data might be measures, the rule to follow is: if a number makes sense
when it is aggregated, then it is a measure. For example, it makes sense to aggregate daily
volume to month, quarter and year. On the other hand, aggregating names or addresses
would not make sense; therefore, names and addresses are not measures. Typical measures
include volume, sales and cost.

413
 Quantitative Methods in Audit and Control

The hierarchies and the dimensions allow the business analyst to slice and dice the
cube according to its needs during data analysis. The cube draws its power from the fact that
aggregation (total sales per quarter, per total market, per total client etc) is already
calculated and the queries against the cube are very fast in response.
By multidimensional operations (drill up, drill down, pivoting, filtering) the business analyst
can slice and dice according to his needs for better understanding of data.

4. Problem analysis

Multidimensional models facilitate and allow fast simulations and what-if analysis
through the support they offer in implementing multiple scenarios of ‚best case’ and ‚worst
case’ type.
The paper presents the case of a Romanian steel producing company with
subsidiaries in Galaţi, Iaşi, Hunedoara and which sells steel products to national and
international clients. The applications was developed used Hyperion Planning 9.3.1.
Here there are some cascading questions a steel production company could make
during it’s planning and forecasting process. Let’s suppose the company has in plan to sell
100.000 tones of steel next year.
1) What if instead will have to sell 150.000 tones or 70.000 tones of steel (because
it gains new clients or loses some of its old clients)? What’s the impact on the production
capacity and the workforce? In order to produce more: it would have to make new hires or
buy new assembly lines and ovens? More raw materials would be needed?
2) What if the petrol price will increase? What if the price of the raw materials will
increase? What if the Euro – Dollar rates will increase or decrease? What would be the
impact on the final price for the client?
The analysis dimensions of the steel company are:
• Year: the years of planning and forecasting;
• Period: with months grouped in quarters (budgets and forecasts are made at a
monthly level);
• Scenario: can be Budget, Forecast, Actual (an actual scenario allows comparison
with real actual data and helps define the budget);
• Version: version 0, version 1, version 2, version 3 etc (for managing several versions
of Budget and Forecast till the final approved ones);
• Client: a dimension containing the company’s clients grouped in National clients and
International clients;
• Product: steel pipes, tubes of certain length, thickness, hollow section, weight, shape,
raw material composition.
The sale price of a tone of steel per product has several components:
• Base price: it is the price that represents the cost of the raw materials that
participated in creating 1 tone of product;
• Extra price: it is calculated monthly. For producing stainless steel special metals, with
very dynamic prices are used. These metals are nickel, chromium and molybdenum.
The extra-price reflects the price of these metals on the stock market and it is
influenced by the Euro-Dollar exchange rate.
• Transportation price: is the cost of transportation;
• Other price: other costs in producing steel.

414
 Quantitative Methods in Audit and Control

Figure 6. Product and Client dimensions

The simulation needs are the following:

• Tons: what happens with revenues if the society sells an extra tons amount to a
certain client, for a certain product, on a certain period or a combination of clients,
products and periods?
• Base price: what happens with revenues if the base production price increases or
decreases by a certain amount for a certain client, product, period or a combination
of clients, products and periods?
• Extra price: what happens with revenues if the extra production price increases or
decreases by a certain amount for a certain client, product, period or a combination
of clients, products and periods?
• The dollar effect: how does the Euro-Dollar exchange rate affects the price
components (base, extra, transportation, other) and what’s the effect on the final
revenues?
The selling prices are in Euros. The Euro-Dollar exchange rate affects only certain
clients (external clients who operate in dollars) on all price components. For clients buying in
Euro, the Euro-Dollar exchange rate affects only the extra price.
VAT is calculated only for Europe and simulations on base price and extra price are
done only for products that contain nickel.
The phases for realizing plans, budgets and forecasts are:
A. Based on the historical and new contracts the company has with its clients, the
commercial department performs an initial prediction for tones to sell to the clients
and the selling prices. The prevision is made for tons and price components (base
price, extra price, transport, others) for every combination of clients and products, for
every month of the year.
B. The planning and controlling department, based on the initial prediction received as
input from the commercial department and based on other parameters (the VAT
percent, the Euro-Dollar rate) performs a prediction of the profit and loss account
“sold production”. Based on the payment conditions (the number of days the client
pays), from the profit account “sold production” the balance sheet account “payments
receivables” is calculated. The controlling department also estimates the remaining
profit and loss and balance sheets accounts.

415
 Quantitative Methods in Audit and Control

The profit and loss accounts impact on the balance sheet accounts. Part of the
balance sheet accounts are calculated using automatic routines from the profit and loss
accounts. For example, “payments receivables” can be automatically calculated form the
“sold production” account.
C. The financial department performs simulations and what-if analysis on number of
tones sold, the price components, the VAT percent, the evolution of the Euro-Dollare
exchange rate. The simulations have a direct effect on the profit and loss account
and the patrimonial state of the organization. Simulations are performed in different
simulation versions, any simulation version will display it’s own profit and loss
account and patrimonial state.

5. Multidimensional model design

The prediction of the commercial activity is done at a detail level of year, scenario,
version, subsidiary, product, client. The profit and loss and balance-sheets accounts are
predicted at a level of detail of year, scenario, version and subsidiary. The different level of
granularity for the two activities conduct us to build an application made up of two cubes:
- one cube for predicting the sold production where simulations on tones and prices
are to be executed (cube Tons).
- one cube for the master budget made up of the Profit and Loss account and the
Balance Sheet (cube MstBdg).

The facts for cube Tons are: tons, prices (base price, extra price, transport, other),
sold production, gross revenue, net revenue, client credits.
The facts for cube MstBdg are the accounts in Profit and Loss and Balance Sheet.
The sold production and client credits calculated in cube Tons (granularity: version,
subsidiary, month, client, product) feed with automatic routines the members sold production
and payment receivables in the MstBdg cube (granularity: version, subsidiary, month).

Table 2. Dimensions for cubes Tons and MstBdg

Dimensions in cube Tone Dimensions in cube MstBdg
Measures (account) Measures (account)
Hierarchy „parametri” Hierarchy „Contul de profit si pierdere”
Hierarchy „masuri vanzare” Hierarchy „Bilant contabil”
Year Year
Scenario Scenario
Version Version
Period Period
Entity (subsidiary) Entity (subsidiary)
Client -
Product -

5.1. The Tons cube

It’s goal is to calculate sales value derived from selling tons of different products to
different clients. The calculus is performed at a very detailed level (month, product, client).
The cube allows fast simulations on tons and prices.

416
 Quantitative Methods in Audit and Control

The simulation problem exposed above can be very easy and elegantly solved by
choosing the appropriate structure of the account (measures) dimension. The analyst has to
design the accounts dimension so that he can fully leverage the potential of the
multidimensional calculation engine. In order to satisfy the simulation needs presented
above, the appropriate facts structure is presented in the figure below:

Figure 7. Accounts dimension for cube Tone

Level 0 members are members that don’t have children. Data is usually inserted at
level 0 member combinations.
Aggregated or upper level members are members that have children. The upper
level members are automatically calculated in base of the lower level members and their
consolidation operation. The consolidation operation can be:
(+) Adds the member to the result of previous calculations performed on other members
(is the default operator);
(-) Multiplies the member by -1 and then adds it to the sum of previous calculations
performed on other members;
(*) Multiplies the member by the result of previous calculations performed on other
members;
(/) Divides the member into the result of previous calculations performed on other
members;
(%) Divides the member into the sum of previous calculations performed on other
members. The result is multiplied by 100 to yield a percentage value;

417
 Quantitative Methods in Audit and Control

(~) Does not use the member in the consolidation to its parent.

For example:
(1) pret_baza_efect$ = pret_baza * M_efect$_pret_baza;
and (2) pret_baza = pret_baza_input + pret_baza_delta;

Hierarchy ‘Parametri’ contains:

- ‘TVA%’ indicates the VAT percent (is fixed for a scenario, cannot vary from one
version to another);
- ‘param_zile_client’ indicate the client paying conditions (number of days the client
pays the products he bought) ;
- ‘rata_schimb’: the user saves in this member the Euro-Dollar exchange rate for the
current version of simulation.
- ‘efect_$’: is automatically calculated by the system and indicates the dollar effect
multiplier.

(3) efect_$ = Euro-Dollar exchnge rate of version stasrt / Euro-Dollar exchange

rate in current version.
Hierarchy ‘Masuri_vanzari’ contains members of type:
- input type members (tone_input, pret_baza_input, extra_pret_input, transport, alte
preturi) are filled in the start version (version_0) of the BUDGET scenario with the
budget data provided by the commercial. This data is the basis for simulations.
- delta type members (tone_delta, pret_baza_delta, extra_pret_delta) are used in
simualtions on tones on prices (tons sold increase/decrease by certain tons, base
price and extra price increase and decrease by a certain amount).
- dollar effect multiplier members indicate the effect of the Euro-Dollar exchange rate
on the price component.
- Gross and net revenues.
Any final price component is calculated with the formula:
(4) price cu efect dolar = pret fara efect dolar * multiplicatorul efect dolar;
The account structure forms the net price as follows:
(5) pret net = pret brut – Transport cu efect dolar
= (pret baza cu efect dolar + pret extra cu efect dolar
+ alte preturi cu efect dolar) – Transport cu efect dolar.

Hierarchy ‘Masuri_temporare’ contains members:

- credit grup
- credit terti
that are used in order to calculate client credits resulted from selling products to clients. The
calculation of client credits is done based on the contractual paying conditions (the number
of days after delivery in which the client performs payment), conditions expressed with the
member ‘param_zile_client’.
For simulation, the business analyst could follow the following steps:
Step 1. Start simulation in version_0
The commercial office of the organization provides an initial version of the budget
by inputting into the system:
- Tons and prices (base, extra, others, transportation) for all the combination of
products, clients, geography and months;
- The VAT percent.

418
 Quantitative Methods in Audit and Control

Step 2. Simulations in version_n

- 2.1. In order to keep clear the start data, data in version_0 is copied into version_n.
Multidimensional software solutions contain data copy functionalities.
- 2.2. The business analyst inserts new exchange rates for the simulation version and
the system automatically calculates the dollar effect multiplier.
Step 3. The user simulates tons and price variations in the delta type members by launching
specific business rules for simulations. The business rule are parameterized calculation
scripts that write down values into the delta members
Step 4. The database is calculated by launching a calc script that calculates various revenue
as tons * price.
Multidimensional software contain easy to use scripting languages that allow
implementation of calculations (business rules).
Step 5. Using a front-end tool to interrogate the cube, the user can investigate the result of
the simulations.
Gross revenues on the simulation versions have changed due to changes of the
Euro-Dollar rates, the extra-tons sold and the increase of the base price.

5.2. The Master Budget cube

Its mail goal is to offer support for realizing the predicted Profit and Loss and the Balance
Sheet.
The account dimension contains two hierarchies:
- the Profit and Loss hierarchy accounts (”Contul de profit si pierdere”);
- the Balance Sheet hierarchy accounts (”Bilantul contabil propriu-zis”).

Figure 8. Accounts dimension for cube Master Budget

419
 Quantitative Methods in Audit and Control

The account „Productia vanduta” (sold production) is imported from the Tons cube
with an automatic routine (the corresponding fact is „Venit brut”) at an aggregated level of
total product and total clients.
Account „Creante comerciale interne” and „Creante comerciale externe” are
imported from the Tons cube using automatic routines (the corresponding facts are
„credit_grup” and „credit_terti”) at an aggregated level of total product and total clients.
Very version of the MstBdg cube contains the effect of the simulations performed on
the same version in the Tons cube.

6. Conclusions

Multidimensional cubes are a powerful instruments for analyzing huge volumes of

data. Regardless of the software used to create multidimensional applications (Hyperion
Essbase, IBM Cognos, SAP BW etc), multidimensional analysis draws its power from the
following elements:
- The possibility to easily analyze the data by slicing and dicing;
- Calculations are easy to implement;
- Aggregated data is already pre-calculated (that implies a fast response time for
queries).
Multidimensional models are used both in analyzing current data and as a support
in planning and forecasting processes. The application designer has a major role when
designing a multidimensional model for simulations. Designing a multidimensional cube is
more than simple work; it can be seen as art considering that an efficiently designed account
structure solves a lot of the simulation issues (especially calculations) and saves the model
from future re-designing in order to adapt it to new needs and requests.

References

1. Agrawal, R., Gupta, R. and Sarawagi, S. Modeling multidimensional databases, Proc. of 13th
Int. Conf. On Data Engineering (ICDE), IEEE Press, 1997, pp. 232-243
2. Cabibbo, L. and Torlone, R. A logical approach to multidimensional databases, Sixt Int.
Conference on Extending Database Technology (EDBT, ‘98), Springer-Verlag, 1998
3. Cabibbo, L. and Torlone, R. Quering multidimensional databases, SEDB, 1997
4. Caraiani, C. and Dumitrana, M. Contabilitate de gestiune & Control de gestiune, Ed.
InfoMega, Bucharest, 2005
5. Codd, E.F., Codd, S.B. and Salley, C.T. Providing OLAP (On-line Analytical Processing) to
User-Analysts: An IT Mandate, Associates Arbor Software, 1993
6. Feleaga, L. and Feleaga, N. Contabilitate financiara. O abordare europeana si
internationala, editia a doua, Ed. Economica, Bucharest, 2007
7. Gyssens, M. and Lakshmanan, L.V.S. A foundation for multi-dimensional databases, VLDB,
Athens, Greece, 1997
8. Ivanova, A. and Rachev, B. Multidimensional models-Constructing DATA CUBE, International
Conference on Computer Systems and Techonologies – CompSysTech, 2004
9. Kerbalek, I. (coord.) Economia intreprinderii, Ed. Gruber, Bucharest, 2007
10. Luecke, R. Manager’s Toolkit: The 13 skills managers need to succeed, Harvard Business
School Press, Boston, Massachusetts, 2004

420
 Quantitative Methods in Audit and Control

11. Sapia, C., Blaschka, M. Höfling, G. and Dinter, B. Extending the E/R model for the
multidimensional paradigm, Advances in database technologies, LNCS Vol 1552,
Springer-Verlag, 1999

1
Daniela Ioana Sandu
Thoroughgoing studies in Informatics Systems for Management (graduated 2002) and bachelor degree in
Economics Informatics (graduated 2001) at the Faculty of Economic Cybernetics, Statistics and Informatics,
University of Economics, Bucharest, Daniela Ioana Sandu is currently a PhD candidate in the field of Economic
Informatics at University of Economics and a Senior Business Intelligence Consultant at IBM-Italy. Her interests are
in data-warehouses, business intelligence, business performance management, financial management.

2
Codification of references:
Agrawal, R., Gupta, R. and Sarawagi, S. Modeling multidimensional databases, Proc. of 13th Int. Conf.
[1]
On Data Engineering (ICDE), IEEE Press, 1997, pp. 232-243
[2] Cabibbo, L. and Torlone, R. Quering multidimensional databases, SEDB, 1997
Cabibbo, L. and Torlone, R. A logical approach to multidimensional databases, Sixt Int. Conference on
[3]
Extending Database Technology (EDBT, ‘98), Springer-Verlag, 1998
Codd, E.F., Codd, S.B. and Salley, C.T. Providing OLAP (On-line Analytical Processing) to User-
[4]
Analysts: An IT Mandate, Associates Arbor Software, 1993
Gyssens, M. and Lakshmanan, L.V.S. A foundation for multi-dimensional databases, VLDB, Athens,
[5]
Greece, 1997
Ivanova, A. and Rachev, B. Multidimensional models-Constructing DATA CUBE, International
[6]
Conference on Computer Systems and Techonologies – CompSysTech, 2004
Sapia, C., Blaschka, M. Höfling, G. and Dinter, B. Extending the E/R model for the multidimensional
[7]
paradigm, Advances in database technologies, LNCS Vol 1552, Springer-Verlag, 1999
Luecke, R. Manager’s Toolkit: The 13 skills managers need to succeed, Harvard Business School Press,
[8]
Boston, Massachusetts, 2004
[9] Kerbalek, I. (coord.) Economia intreprinderii, Ed. Gruber, Bucharest, 2007
Caraiani, C. and Dumitrana, M. Contabilitate de gestiune & Control de gestiune, Ed. InfoMega,
[10]
Bucharest, 2005
Feleaga, L. and Feleaga, N. Contabilitate financiara. O abordare europeana si internationala, editia
[11]
a doua, Ed. Economica, Bucharest, 2007

421
 Quantitative Methods in Audit and Control

STANDARDS REVIEW ON MISSION OF MANAGEMENT

INFORMATION SYSTEMS AUDIT

Delia BABEANU
Ph.D, Information System Supervisor, Lecturer,
Faculty of Accounting and Management Information Systems,
University of Economics, Bucharest, Romania

E-mail: [email protected] Web page: www.cig.ase.ro

Valerica MARES
Ph.D, Lecturer, Faculty of Accounting and Management Information Systems,
University of Economics, Bucharest, Romania

E-mail: [email protected] Web page: www.cig.ase.ro

Abstract: The purpose of auditing is to verify that all hardware and software functions,
automated processes, declared and published performance criteria are producing correct
results, within the confines of system integrity, security measures and other control
mechanisms, in accordance with functionality, originally envisaged, designed or modified.

Key words: standards for information systems audit; risks management; information security;
IT governance

Introduction

The scope of an information systems audit may be as far and wide to cover the entire
lifecycle of the technology under scrutiny, including the correctness of computer calculations,
or a basic test to verify compliance with a specific declared objective. The "scope" is of an
audit is dependent on its declared objective, decided upon from the outset.
Audits may be initiated as a result of some concern over the management of assets.
The concerned party may be a regulatory agency, an asset owner, or any stakeholder in the
operation of the systems environment, including systems managers themselves. Each and
every party may probably have an objective in initiating and commissioning the audit. Such
an objective may be to validate the correctness of the systems performance or calculations,
confirming that systems are appropriately accounted for as assets, to assess the operational
integrity of a series of automated processes, to verify that confidential data is not
compromised by being exposed to unauthorized persons, or it may even be a multifaceted
combinations of the above mentioned aspects in addition to a wider ranging information
systems issues of lesser or greater significance to an organisation, which by its very nature,
may vary from one place to another. Selected various objectives of an audit will ultimately
determine its scope.
The purpose of auditing is to verify if all hardware performances are used according
to the software ones, at designed parameters. In order to achieve these normal working

422
 Quantitative Methods in Audit and Control

parameters of computer networks have been defined as well as those peripheral devices.
It is important that the audit starts from the results of a previous audit of the
company. The existing documents, created by a previous mission, should be analyzed, after
which all the subsequent changes to the system will be verified. If these existing documents,
as well as the documentation for further changes satisfy the need for information of the
auditor, he will proceed to controlling the implementation of the changes.
The time period in which the audit takes place has to be well defined. Collecting
samples is done using files that keep the history of the network, user rights, and hardware
and software resources.
The audit of information systems is not different from other audits; it consists of the
analysis of the systems referring to an activity of the company. In this sense, it is required to
define information applications that represent an integrated set of programs, data and
administrative procedures. Examples of such applications are: primary accounting
applications, salary payment report applications, application for managing stocks, etc. The
largest part of information applications are considered processes articulated around various
stages like entries, processing, data storing and obtaining results (Nastase, 2007).

Standards presentation

The performed audit is based on current laws, standards and norms. One of these
is standards series 27000. Standards that can be applied and are part of this series refer to:
The family of standards for SMSI – Information Security Management System
(ISO27000 – ISO27010, http://www.iso27001security.com/html/27000.html) which covers
the specifications of the system, measurements, an implementation guide, an audit guide
and the management of risks.
The following are part of this category:
• ISO 27000 – fundamental elements and vocabulary (completed at the end of 2008)
which:
9 explain the terminology for all the series of standards 27000 (marketing)
9 explain basic principles and definitions that vary from one country to another
9 these principles will have an impact on other standards as COBIT (IT processes) and
ITIL (Providing IT services – Service Delivery) and eliminates all confusions
• ISO 27001 – requirements of a SMSI – Certification Process (is based on ISO 27002)
9 -certifying SMSI – published in November 2005 and operational on January 30 2006
(www.iso27001certificates.com);
9 -classification/improvement of the requirements of the PDCA process
(http://27001.denialinfo.com/pdca.htm), which covers:
o -the scope of SMSI (figure 2), evaluating risks, selecting controls, appliance
declaration, reviewing risks, SMSI internal audit, real results and
measurements, plan for treating risks and controls
• ISO 27002 – Good practice code for managing informational systems:
9 it has 11 sections which treat the protection of informational assets (it was published
in April 2007)
9 -133 detailed controls (based on the process of evaluating risks and the business
environment)

423
 Quantitative Methods in Audit and Control

9 -covers outsourcing purchasing and delivery services, current issues and

management issues, security services at employment and during a contract of an
employee, a guide for risk management and managing incidents, mobile remote or
distributed communications,
• ISO 27003 – SMSI Implementation Guide (will be available in 2009)
9 Implementing the guide that will provide support for the new requirements of the
standard
Annex B of BS7799 Standard - The second part has the following stages: overview,
the responsibilities of the management, conformity with governance and rules,
human resources and personnel security, managing assets, availability/continuity of
business processes, managing informational incidents, access control, case studies
for risk management (http://17799.standardsdirect.org/iso-17799.htm)
9 Implementing a PDCA implies identifying assets, identifying threats, evaluating and
treating risks, analyzing and improving controls.
• ISO 27004 Metrics and measurability of SMSI (at the end of 2008). The objectives of
this standard are:
9 a real evaluation of SI controls and objectives
9 a real evaluation of a SMSI
9 offers indicators for management assistance
9 improving SI facilities
9 provides entries for SI audit
9 real communication at the information systems management level
9 input the process of risk management
9 output for internal comparisons and benchmarks (i.e. measuring controls and
processes performance)

Figure 1. SMSI Planning stages

424
 Quantitative Methods in Audit and Control

• ISO 27005 SMSI risk management (end of 2008)

9 -a new risk management standard for information security
9 -risk analysis, evaluating risks from informational security (identifying assets, threats
and vulnerabilities
9 -treating informational security risks (presented in figure 1)
Annex a – goal
Annex b – identifying and evaluating assets
Annex c – common vulnerabilities (http://www.27001.com/catalog/7)

Figure 2. Risks treatment of information security

(Source: draft ISO 27005)

• ISO 27006 – SMSI accreditation guide (certification contents)

9 necessary for increasing rigors and underlining the contents of certification which is
required by the organization (business needs, communications and practice);
9 Operational from January 2007;
9 General requirements (impartiality guide);
9 Organizational structure applying ISO/IEC 17021;
9 Resource requirements: managerial competence, subcontracts;
9 Informational requirements – guiding certification results;
9 Process requirements – guiding SMSI audit.
• ISO 27007 – SMSI auditing guide (from 2009)
9 Guide for auditing and SMI auditing content certification accreditation. This family of
standards is represented in figure 3: Standards family applicable to a SMSI

425
 Quantitative Methods in Audit and Control

Figure 3. Standards family applicable to a SMSI

9 Specific requirements for certain sectors of the economy (ISO 27011-ISO27030) –
Telecom (global) ISO 27011, Health (UK) ISO 27799; Automotive (Germany; Korea;
Sweden); Lottery at international level.
9 Operational guide (ISO27031 – ISO27059) for which the publication date has not
yet been decided. This series contains:
9 -ISO 27031 ICT standard on business continuity
9 -ISO 27032 cyber security
9 -ISO 27033 - network security.
9 -ISO 27034 - application security.
The pursued objectives can be found in the following table:

Table 1. Informational systems audit objectives

Major objectives Implementing a good practice
Evaluating existing or replaceable controls
Configuring key points for information security
Reducing frequency/impact of major incidents
Important objectives Aligning to the internal security policy
Integrating in the risk management program
Identification of specific requirements for a certain activity domain
Increasing existing investments
Other objectives Increasing competition advantages
Identification of requirements at the industrial branch level
Answering a pressure by a third party
Obtaining a minimum cost

Audit materiality, covered in S12 standard, consists of basic principles and essential
procedures clearly identify, which are mandatory together with the guide for the elaboration
of these procedures.

426
 Quantitative Methods in Audit and Control

Figure 4. Strategic objectives of the audit

(Adapted for SIG after http://imm.protectiamuncii.ro and http://hwi.osha.europa.eu)

The relevance of materiality consists in the quality of information in a system which

a society requires in order to publish all significant information. The materiality threshold is
the one from which risks are important in evaluating a society.
This valuation can be done from a quantitative or qualitative point of view and in
certain cases (an informational system permanently exposed to risks) a combination of the
two methods.
The evaluation of the materiality is a judgment problem and includes
considerations over the effect of organization’s ability to achieve objectives in events like
errors, omissions, irregularities or illegal acts which could substantially modify the results of
the controls of threats in the audited area. When evaluating the materiality one has to take
into account the errors accepted by the management, by the SI auditor, by the objectives
assigned to the system or financial transactions processes, stored information, hardware,
architecture and software, infrastructure network. Operating system, development and
testing environment (www.isaca.org).
Examples of measures for evaluating the materiality:
• critical business processes supported by the system or applications (data acquiring,
processing, reporting etc.)
• databases with critical information from the system or operations
• number and type of developed applications
• number of users that access the informational system
• number of managers, directors that work with classifies information according to
their privileges
• critical network communication from within the system or operations
• system or operations cost (hardware, software, personnel, outsourced services, alone
or in combinations)

427
 Quantitative Methods in Audit and Control

• potential cost of errors (in terms of sales losses, lost guarantee, uncovered
development costs, advertising cost required by guarantee, rectifying costs, health
and safety, unnecessary production costs, etc)
• number of transactions requested over a period of time
• nature and quantity of manipulative materials
• requirements related to service contracts and costs of penalties
• Other penalties
Reporting materiality supposes determining findings, conclusions and
recommendations that are to be reported. Weaknesses control should be considered
materiality and reported if the absence of control causes errors in ensuring objectives
controls.

Conclusion

Support information and processes, facilities, computer networks and the

connection between them are the most important assets of a business. In order to manage
these assets and to have business continuities, one has to implement SMSI standards in
every company.
We propose transforming the components of the informational system and the
information system in values and establishing a threshold for materiality based on value,
computed in the respective national currency, which could be taken as the theory of
materiality transformed in significance threshold, as is the case of the financial accounting
audit.

References

1. Nastase, P., Eden, A., Nastase, F., Stanciu, V., Popescu, G., Gheorghe, M., Babeanu, D., Gavrila,
A. and Boldeanu, D. Auditul si controlul sistemelor informationale, Ed. Economica,
Bucharest, 2007, pp. 17-22
2. * * * http://www.isaca.org
3. * * * http://hwi.osha.europa.eu
4. * * * http://27001.denialinfo.com/pdca.htm
5. * * * http://www.iso27001security.com/html/27000.html
6. * * * http://www.27000.org/
7. * * * http://www.27001.com/catalog/7
8. * * * http://17799.standardsdirect.org/iso-17799.htm
9. * * * http://www.iso27001certificates.com

428
 Quantitative Methods in Audit and Control

A MODEL OF CREDIT BUREAU IN SERBIA – INSTRUMENT FOR

PRESERVING STABILITY OF THE BANKING SECTOR IN
CONDITIONS OF THE GLOBAL ECONOMIC CRISIS

Vladimir SIMOVIC1
PhD Candidate, Faculty of Organizational Sciences
University of Belgrade, Serbia

E-mail: [email protected]

Vojkan VASKOVIC2
PhD, Assistant Professor, Technical Faculty Bor
University of Belgrade, Serbia

E-mail: [email protected]

Dusan POZNANOVIC3
MSc, Public Procurement Agency, Belgrade, Serbia

E-mail: [email protected]

Abstract: This paper presents the characteristics of the banking system in serbia before and
during the global financial crisis. The model of the credit bureau in serbia which, according to
its technical characteristics and the level of business performance, represents the original
solution is analyzed. Its implementation, in conjunction with other control mechanisms, has
provided the stability of the banking sector in terms of crisis. Consequently, the control of
liquidity in the banking sector is achieved as well as the control of the expansion of credit
activities, with the maintenance of population and economy indebtedness at optimal level,
which is of great importance in terms of global crisis when economic policy makers in serbia,
faced with a pronounced deficit in balance of payments of the country, as one of economic
policy measures aimed at improving the balance of payment position, implement the measure
of controlled reduction of private demand.

Key words: credit bureau; financial crisis; liquidity risk; loan placement; Serbia

1. Introduction

World economic crisis has a direct impact on the countries that dominate the
international capital flows and international trade. Unlike them, the developing countries
and countries undergoing the transition process are indirectly influenced by the crisis which,
in the context of the impact on the financial sector, is manifested as a crisis of liquidity and
difficulties in development and reform of financial institutions (Vjetrov A. et al. 2009). Serbia
belongs to this group of countries.

429
 Quantitative Methods in Audit and Control

Manifestation of the crisis of liquidity brings into question the stability of the
banking system in Serbia, as the holder of the entire financial system. Our intention is to
point out on the credit bureau as one of the instruments for preserving liquidity in the
banking sector in Serbia as well as the stability of the same, before and during the crisis.
Information that are analyzed in this paper were collected on the basis of research
conducted in the Association of Serbian Banks (ASB) under whose authority functions the first
credit bureau in modern Serbian history.
The paper is organized as follows. Section 1 is the Introduction to the theme.
Section 2 represents an overview of the characteristics of the banking system in Serbia in the
period before and during the global financial crisis, with special emphasis on measures that
the National Bank of Serbia (NBS) has undertaken with the aim of preserving the stability of
the banking sector in terms of global crisis and shaken confidence of the Serbian citizens in
it. Section 3 represents an overview of the functional characteristics of Serbian credit bureau
model as well as its specificity which have caused it to be ranked as one of the best in the
World by the World bank. Section 4 points out the importance of credit bureau in Serbia as
an instrument for the preservation of liquidity in the banking sector.

2. Characteristics of the banking sector in Serbia before and during

the World economic crisis

Financial system in Serbia is a network of institutions which consists of 34 banks, 22

insurance companies, 17 leasing companies and 9 voluntary pension funds. Within the
financial system of Serbia banks have a very dominant role and run with 90% of total
financial assets.
National Bank of Serbia carried out a radical reform of the banking sector during
2001 and 2002 which resulted in the closure of 23 insolvent banks, thus erasing almost 70%
of all assets of the banking sector (Ostojic S. 2002). Reform of the banking sector led to a
reduction in the number of banks in which the state is the majority owner and increased the
number of banks with foreign ownership whose arrival has increased competition in the
market and the efficiency of the banking sector. In today's conditions in the balance sheet
amount of the banking system in Serbia, the dominant share of 75% are foreign-owned
bank.
Since 2004 private sector loans in Serbia recorded dynamic growth which is a
consequence of the low starting base and the fact that before 2003 private sector loans
nearly were not approved at all. The main source of credit activities was the growing deposit
potential, but significant funds were also provided by means of recapitalization of banks and
loans from abroad. The trend of increasing dependence on loans from abroad manifested in
2005, began to drop in the second half of 2006. Consequently, in today's conditions, 7.6%
of sources of funds of banks in Serbia are those loans (National bank of Serbia, 2008).
The growth of the banking sector in Serbia in 2008 was slowed as a result of
restrictive monetary and prudential policy, and in the fourth quarter as a result of the global
economic crisis. The period of credit expansion that lasted from 2004 ended in the last
quarter of 2008. Since then credit activity achieved the minimum real growth in some
months, while retail loans achieved a nominal decrease. This situation is caused by the
reduction of demand for loans due to negative macroeconomic trends as well as by the
reduction in credit supply due to the bad situation in terms of sources of liquidity and

430
 Quantitative Methods in Audit and Control

minimized tendency towards risk by the banks. Retail loans recorded a growth of 20% in
2008, which is significantly below the 54% of real growth recorded in 2007. The decline in
retail loans started during the 2007 thanks to the measures of the central bank issued in
terms of expansion of retail loans. The fact that retail loans declined gradually is important
for preserving financial stability in conditions of financial crisis. The decline of credit activity is
mostly pronounced for the retail and cash loans to which the prudential measures of the
NBS were targeted in 2007.

Cash loans
Consumer loans
Housing loans
Credit cards
Number of credit
users

Figure 1. The growth of retail loans by purpose

National Bank of Serbia did not react only to the consequences of the crisis, but a
responsible monetary policy and prudential measures in the time before the crisis acted
preventively, alleviating the negative effects of the global financial crisis. In this sense, the
NBS has implemented a restrictive monetary policy (high key policy rate and withdrawal of
excess liquidity) prudential (comprehensive and conservative risk weights, reducing exposure
to foreign currency risk, limiting the indebtedness of the population) and administrative
measures (high required reserve on foreign currency savings and loans from abroad, limiting
the relationship of gross retail placement and capital), and tightened control of commercial
banks and also established the first private credit bureau.
In terms of the financial crisis in Serbia it turned out that the greatest vulnerability
of the domestic financial system is high share of indexed loans (70%) which increased the
foreign exchange and interest rate risk. The fact that the real sector (due to a low share of
exports in GDP) and population (due to income in dinars) are exposed to foreign exchange
risk is having the great impact on the fact that nominal depreciation leads to increase of
defaults and impairs the quality of assets of the banking system.
One of the first effects of the global financial crisis in Serbia, manifested through
the withdrawal of deposits from banks. Bad experience from the past, in terms of savings
"trapped" in the pyramidal and some state banks, had a negative psychological impact on
depositors in Serbia who widely started to withdraw deposits from banks, which have had a
short-term negative effect on banks' foreign currency liquidity. In order to restore the shaken
confidence of citizens in the banking sector, Serbian Government adopted a set of measures
which could be systematized as follows:
• The state guarantee for savings was increased from € 3000 to € 50,000 per
depositor if the bank went into bankruptcy. The decision to increase the amount of
insured deposit was made on the proposal of the European Commission.

431
 Quantitative Methods in Audit and Control

• In order to encourage savings, starting from January 1st 2009, Serbian Government
temporary abolished income tax for foreign currency savings, which amounted to
20%. In 2010 this tax will be charged at a rate of 10%.
• Temporarily, until the end of the 2012, capital gains tax (20%) was abolished as well
as the tax on the transfer of absolute rights (0.35%) realized through securities
trading.
In addition to the Serbian Government, NBS has prepared a set of measures to
mitigate the negative effects of the global crisis on the financial sector in Serbia:
• The supervision of the financial system through intensive control of daily liquidity,
deposits and foreign currency reserve of banks was reinforced. A new regulatory
framework that enables regular data collection on uncollectable receivables was
adopted. In addition, the collection and exchange of information on the financial
conditions of centrals of the banks that operate in the country was improved and
control of financial accounts on a daily basis was reinforced.
• Required reserves for funds taken in overseas aren't being calculated retroactively
from October 1st 2009 for bank loans from abroad (untill then the required reserve
was 45%), subordinated capital from abroad (20%) and borrowing of th e financial
leasing companies (20%)
• National Bank of Serbia has ordered banks to change the structure of the required
reserves which are held on the account with the NBS, so that instead of the former
90% of reserves in foreign currency and 10% in dinars, now required reserves consist
of 80% of reserves in foreign currency and 20% in dinars.
• Reduction of penalty interest rate for dinars from 31.75% to 23.63% and for foreign
currency from 31,75% to three-month Euribor plus 10%
Thanks to the recapitalization of banks, as well as restrictive prudential and
monetary policy of central bank, the banking sector in Serbia welcomed the spread of
financial crisis with a high degree of resistance. Unlike banks in Europe and America, the
Serbian banking sector is well prepared for external challenges. For example, depositors in
Serbia are provided with a high level of protection through high required reserve in the
amount of 40% for the new foreign currency savings. High capital adequacy of 28.1%
(among the highest in Europe) and low dependence on bank loans from abroad, as well as a
wide deposit base are mitigating the effects of the liquidity shock. Also, a third of the
balance sheet sum of the banking sector are cash, deposits with NBS and securities of the
central bank. The fact that local deposits are over 70% of total liabilities confirms the stable
structure of sources of funds. Observed by sectors, the domestic deposit base as the primary
source of financing for banks is made of population deposits (49% of total deposits).
Favorable indicator from the point of the compliance of sources of financing and loans is the
fact that the total retail loans are almost completely covered by term deposits of the
population, and that ratio is much better than in many European countries. From the point of
maturity of term deposits inconsistency between sources of financing and loans is evident -
long-term loans exceed the long-term sources of financing multiple times. Thanks to the
policy of required reserves on borrowing abroad and the domestic foreign currency deposits,
the banki ng sector in Serbia is characterized by coverage of deposits in foreign exchange
reserves amounting to 86.3% which is much more than this rate in most countries of the
region in which it amounts on average 35%.

432
 Quantitative Methods in Audit and Control

Considering the retail loans to GDP the population of Serbia is one of the least
indebted in the region. In Serbia, Macedonia and Albania approved retail loans are 12-15%
of GDP, while in Romania, Hungary and Bulgaria, that ratio is 20-25% of GDP. In Croatia,
this percentage is up to 40%. From the point of indebtedness per capita which currently
amounts to € 637, Serbia is in far more favorable position in comparison to some countries
in the region in which the indebtedness per capita is up to € 3750 (Croatia). This is a direct
consequent of the measures taken by NBS by which instalment of the loan can be up to 30%
of monthly income and the introduction of the credit bureau. The expansion of credit activity
which is not accompanied by adequate control mechanisms can jeopardize the entire
banking system through the emergence of liquidity crisis. The introduction of credit bureau in
Serbia aimed to enable optimal alocation of resources based on reliable information for the
creditors and in conjunction with measures of NBS to ensure stability of the banking system,
which is particularly important in terms of the global economic crisis.

3. Serbian credit bureau model

The model of Serbian credit bureau is the result of observed experiences of models
that are more or less successfully applied in practice of a large number of countries. With
certain modifications, solutions which were estimated to have been good, were accepted
and relying on own resources a model, which for most of its functional characteristics
represents the original solution, was developed. The ultimate result is a model of the credit
bureau which was ranked by the World Bank as the as one of the best in the World. Credit
Bureau in Serbia has started its operative work on 2004 as the result of the initiative of the
Association of Serbian Banks with the consent of the National Bank of Serbia and the
Ministry of Finance and Economics. It should be noted that the institution of the credit
bureau has a long tradition in Serbia, since the Inforamtional Credit Department existed in
the Kingdom of Serbs, Croats and Slovenes from 1929, but after the Second World War the
whole system closed its opeations (Vaskovic V, 2007).
The basic assumption, and also the biggest advantage of the Serbian credit bureau
model is the fact that the creditors (72 members of the credit bureau in Serbia) are solely
responsible for the accuracy of the data shown in the credit bureau reports. This fact is
ensured thanks to the unique technological process of the credit bureau in Serbia (Figure 2).
The central database is located in the credit bureau, and the creditors (banks and other
financial institutions) have their own part within the central database. Therefore, banks and
other creditors practically rent private space within the information system of the credit
bureau and are responsible for its maintenance and the accuracy of data in it.
In the banks and with the other creditors, once a day, usually before the end of
business hours, procedures that draw data from their production database (on the credit
activities of their clients) are initiated. As a result the documents for the exchange of data
with the information system of the credit bureau are generated. Those documents are in XML
format, digitally signed and encrypted. It is necessary that the data, which are to be imported
by creditors in the rented private database, meet the criteria of validation. Those criteria use
the logic, syntax and semantic rules for data filtering and only if all rules are satisfied the
data can be stored in the database. This ensures a high degree of accuracy of data that
would later appear in the credit bureau reports. Each creditor namely, the authorized person
in this institution, can only access its private database, and the credit bureau can access the

433
 Quantitative Methods in Audit and Control

private databases of the creditors based only on written consent of the client by which the
report is withdrawn from the credit bureau.

Bank 3

Checker

Private base
Bank 3

D
a
t
C a C
h Private h
e Private base e
Creditor 1 c base Credit bureau c Bank 2
Data Data Bank 2
k Creditor 1 k
e D e
r a r
t
a

R C
Private base
E O
Bank 1
P N
O S
R E
T N
Checker
T

Bank 1

Figure 2. Technological process of credit bureau in Serbia

By the decision of the National Bank of Serbia, which is in accordance with

Directive of European Parliament and The Council for Harmonization of Regulations of the
Member States, it is regulated that before the loan contract is concluded, the creditor
evaluates with all means available, whether the credit applicant regularly pays its obligations
under previously approved loans. This means that the creditor is obliged to withdraw the
report from the credit bureau. On withdrawing the credit report from credit bureau, the
creditor needs to forward to the credit bureau the written consent of a natural person or
legal entity for which he asks for report. Then, the data from private databases of all
creditors are matched and loaded into the reporting database. The data become accessible
and the information is visible in the credit report.
The good side of this model is the fact that the credit bureau can not change in any
way the data stored in creditors private databases. Precisely for this reason, all the
responsibility for the accuracy of data shown in credit bureau reports is fully transferred to
the creditors, which from the credit bureau's point of view significantly facilitates the
resolution of eventual complaints. In the case of the complaints on the accuracy of the data
shown in credit bureau reports by the end users of credit lines, the request is forwarded
directly to the creditor which imported the original data in the rented private space within
the information system of the credit bureau. The creditor is obliged to check the

434
 Quantitative Methods in Audit and Control

reasonableness of the complaint request and if necessary to make the changes of the data
and to forward the complaint request to the credit bureau that monitors its resolution.
In current credit bureau business practice in Serbia on total number of issued
reports (8,501,055) there has been 22.147 reported complaints, which is 0,26 % of the gross
number of issued reports. This data clearly confirms the fact that creditors in Serbia make
their credit decisions based on reliable and high quality information.

Table 1. The number of complaints in relation to the number of issued reports

NUMBER OF
YEAR ISSUED REPORTS PERCENTAGE
COMPLAINTS
2004 65.206 0 0
2005 1.106.725 1.037 0.09
2006 1.710.999 3.389 0.20
2007 2.641.295 7.220 0.27
2008 2.976.830 10.501 0.35
SUM 8.501.055 22.147 0.26

Growth of the number of complaints is in correlation with the number of issued

reports, but still is very low as a result of organizational and technological sophistication of
the model of credit bureau in Serbia which by applying data validation rules and eliminating
the possibility of errors during the process of data manipulation by the credit bureau
provides a high degree of accuracy of data shown in credit reports.
Based on credit bureau report the creditor (bank) decides whether the applicant
would be provided with the requested service. The adequate interpretation of the data
shown in credit bureau reports is necessary for making the optimal decision by the creditor
that minimizes the risk and simultaneously increase the number of users of its services. A
model of credit scoring was developed for this purposes and currently it is used only for
natural persons. This model is calculating credit score by taking the following factors into
account:
1. Orderliness, and/or disorderliness in the discharge of obligations is the first and
dominant factor that influences summary score (35% of total score)
• Debit rate is a second factor according to the influence on the summary score and it
is 30% of total score
• Third factor is the time necessary for settling of irregularities in the discharge of
obligations. The influence of this factor on the overall score is 15%.
• The number of drawn reports from the banks in the past 30 days is the fourth factor,
and it influences the score in 10%.
• The length of time for the use of bank services which influences to a certain extent
on the overall score in the amount of 5 % is the fifth factor.
• Sixth factor is the number of used bank services and it influences the overall score
with 5 %.
One of the essential characteristics of the credit bureau model in Serbia is the fact
that the counters of banks and other creditors are used as branches of credit bureau. This
organization is possible thanks to the fact that for each charged fee for issuing report from
the credit bureau, bank receives 40% of the total sum and the fact that banks and other
creditors are responsible for the accuracy of the data shown in credit reports, so that in terms
of resolving complaints, the most optimal solution is solving it in a direct contact of client and
creditor. Thus, the significant operational expenses for establishing own branches on the
territory of Republic of Serbia have been eliminated.

435
 Quantitative Methods in Audit and Control

4. Credit Bureau - instrument of securing liquidity

in the banking sector

The stability of the banking sector is one of the basic prerequisites of stability of the
entire financial system, particularly in transitional economies in which banks have a
dominant role in the financial system. The stability of the banking sector, among other things
is significantly influenced by different categories of financial risks - credit risk, liquidity risk
and market risks. For the purposes of this paper, special attention will be devoted to the
issue of preserving the liquidity of the banking system due to the direct impact of the credit
bureau on liquidity at bank level, through the control of loan placement. Maintaining
liquidity of the banking system is a complex problem that involves coordinated action of
central bank and individual banks.
One of the major problems in the operation of commercial banks and other
creditors is maintenance of liquidity through monitoring financial discipline of bank clients.
Liquidity represents the ability of the bank to have, at any time, adequate amount of funds
necessary to finance the growth of assets and timely cover all the obligations that are due.
Liquidity risk is the risk of emergence of negative effects on the financial result and capital of
banks due to the inability of banks to meet their outstanding obligations. Liquidity risk is one
of the leading financial risks in the banking sector, whose sources are often other financial
risks like credit and market risks.
Problems with the liquidity of a single bank may have a significant impact on the
banking sector and financial system as a whole. Current crisis points out to the great
importance that liquidity of bank and the banking system has on the overall financial system
and economy both on national and international level.
Liquidity in the banking system in Serbia in 2009 can be described as satisfactory.
The average monthly indicator of the overall liquidity of the banking sector in March 2009 is
1.88, which can be considered a satisfactory level due to the regulatory minimum of 1.
Credit bureau has an indirect impact on the liquidity of the banking sector by
reducing the credit risk which, as already mentioned, can lead to the appearance of liquidity
risk. The introduction of credit bureau institution in Serbia is aimed to reduce the risk of loan
placement, which in combination with other factors resulted in maintaining liquidity in the
banking sector at the optimal level.

0
2003 2004 2005 2006 2007 2008 III 2009

Figure 3. The average monthly indicator of the overall liquidity of the banking
sector in Serbia
Data source: National bank of Serbia (Narodna banka Srbije)

The impact of the credit bureau on liquidity movement is difficult to quantify due to
the fact that the liquidity is conditioned by many factors. Credit Bureau aims to provide

436
 Quantitative Methods in Audit and Control

reliable information that would enable banks to make the adequate credit decision that
allows the optimal allocation of resources and minimize the credit risk. Banks use the
information on the behavior of their clients in the past, so they can predict their future
behavior. The research conducted in banking sector of 34 national economies (Miller M.J,
2003) where the credit bureau operates showed that more than half of the respondents
thought that the posibility of using credit information obtained from the credit bureau for
making the credit decision, makes the time of loan approval shorter, lowers the costs and the
default rate for more than 25%. Impact of the credit bureau on reducing the credit risk, and
indirectly the risk of liquidity will depend on the quality of information which are available to
the users of its services. Depending on the chosen business concept, the credit bureaus can
record positive and/or negative information, about the credit activity of users of the credit
lines. The research conducted by the International Finance Corporatioin (International
Finance Corporation, 2006) suggests that if the credit bureau records both positive and
negative information about the credit users activity, the default rate would decrease by 43%
compared to the situation when making credit decisions is based solely on the negative
information about credit users past behavior. Credit Bureau in Serbia keeps records in its
database containing both positive and negative information on credit history of loan
applicant. Due to this fact and the high quality of available information (which is confirmed
by the low percentage share of complaints in the total number of issued reports) creditors in
Serbia are able to realistically assess the credit risk.
For the purpose of this analysis we would point out the importance of non
performing loans and the impact that the introduction of the credit bureau institution had on
reduction of percentage share of non performing loans in the total amount of approved
loans. According to the methodology of the IMF a loan is nonperforming when payments of
interest and principal are past due by 90 days or more (Svartzman I, 2003). Non performing
loans are significant due to the fact that their greater percentage share in total loans
approved, leads to the reduction of the liquidity of banks, risking to jeopardize the entire
banking system. The banks are pushing their efforts to make an optimal credit decisions,
based on reliable information obtained from credit bureau, in order to reduce the credit risk
and consequently the share of non performing loans in the total amount of approved loans.
Institution of credit bureau, besides a direct impact on reducing the credit risk, indirectly
affects the reduction of the percentage share of non performing loans in the total amount of
approved loans, by increasing the financial discipline of the credit lines users. Figure 4 is a
preview of percentage share of non performing loans in Serbia in the period 2003-2009.

25

20

15
(%)
10

0
2003 2004 2005 2006 2007 2008 III 2009

Figure 4. Non performing loans in Serbia in the period 2003-2009

Data source: National bank of Serbia (Narodna banka Srbije)

437
 Quantitative Methods in Audit and Control

What is evident from Figure 4 is the fact that the introduction of credit bureau in
Serbia had a direct impact on reduction of share of non performing loans in the total amount
of approved loans from 23.8% (2005) to 4.1% (2006). Namely, the credit bureau in Serbia
has started its operative work on collecting and processing data for natural persons in 2004,
and the collection of data for legal enteties started in April 2006. For this reason, the
reduction of share of non performing loans in the total amount of approved loans is evident
only in 2006 and after, as a direct consequence of the introduction of credit bureau. In this
way, the credit bureau confirmed its role as an instrument for the preservation of liquidity in
the banking sector in Serbia and in the end the stability of the same, which is particularly
important in terms of the World economic crisis. The results, shown in Figure 4, confirm the
general expectations that the repayment ability of credit customers would decrease during
recession, since the movement of non-performing loans is affected not only by the control
mechanisms but also the general market trends.

5. Conclusions

This paper points out the problems of the banking sector in Serbia in conditions of
the global economic crisis with special emphasis on measures undertaken by regulatory
authorities in order to mitigate the adverse effects of the same. The special characteristics of
the Serbian credit bureau model are emphasized as well as its unique technical and
technological and organizational structure that makes it an original solution in global terms.
Due to the fact that banks and other creditors are solely responsible for the accuracy of data
shown in credit bureau reports, high quality of data on which creditors base their credit
decisions with a minimized credit risk is ensured. The unique organizational structure of the
credit bureau in Serbia, which implies that the credit bureau uses the counters of banks and
other creditors as own branches in order to reduce operating costs is also described in the
paper.
The hypothesis that the credit bureau is one of the instruments for preserving
liquidity in the banking system, and consequently for ensuring the stability of the same,
which is of especially great importance in conditions of the global economic crisis, is
empirically confirmed in the paper. Analyzing the impact of the credit bureau on the
participation of non performing loans in the total amount of approved loans in the period
from 2003 to 2009, its positive impact on reducing the liquidity risk as the basic assumption
of stability of the banking sector is confirmed.

Literature

1. Miller, M..J. Credit reporting around the globe, Credit Reporting Systems and the International
Economy, MIT Press, Cambridge, 2003
2. Ostojic S. Osvrt na iskustva privatizacije – pouka ili ne, Privredna izgradnja, XLV,1-2, 2002,
pp.35-62
3. Svartzman I. The Treatment of Nonperforming and Interest Accrual, IMF, May 2003
4. Vaskovic, V. Sistemi plaćanja u elektronskom poslovanju, FON, Beograd, 2007
5. Vjetrov A. et al. Svetska ekonomska kriza i posledice po Srbiju, Institut fakulteta za
ekonomiju, finansije i administraciju, Beograd, April 2009

438
 Quantitative Methods in Audit and Control

6. * * * Credit bureau Knowledge Guide, International Finance Corporation, World Bank Group,
Washington DC, USA, 2006
7. * * * Report on the state of financial system, National bank of Serbia, www.nbs.rs, September
2008

1
Vladimir SIMOVIC has graduated on the Faculty of Economics, University of Kragujevac. He holds a Master
diploma in e-Business at Faculty of Organizational Sciences, University of Belgrade. Also, he is a PhD candidate in
the field of e-Business at Faculty of Organizational Sciences, University of Belgrade. His interests are in payment
systems, financial management, SEPA initiative, methodology for the prevention of the credit over-indebtedness.

2
Vojkan VASKOVIC enrolled at the Faculty of Organizational Sciences (computer science) and graduated in 1981.
He was a post- graduate student during 1984/85. He became a Master of Science in 1989. followed by the Ph.D.
in 1997. In the scope of his scientific work up to now, he has published a large number of works, a book (7) and
16 patents out of which 8 are made. He is engaged as a professor on Belgrade Business School. Also, he is
engaged as an Assistant Professor on Technical faculty Bor, University of Belgrade.

3
Dusan POZNANOVIC had finished Electrotechnical Faculty, University of Belgrade. Now Dusan is director of IT
company "Belit d.o.o." and his specialty is managing different projects, creating new technologies in banking and
financial sector. Main projects are: Credit Bureau (national centralized information system for acquisition and
distribution of credit record history for individuals and companies), Creating Banking reporting system for National
Bank of Serbia and Development of Public Procurement Office web portal.

439
 Software Analysis

DIFFERENT SIMULATIONS OF A BILLIARDS GAME

Lucio T. DE PAOLIS1
Assistant Professor, Department of Innovation Engineering,
Salento University, Lecce, Italy

E-mail: [email protected]

Marco PULIMENO2
BSc, ISUFI - Salento University, Lecce, Italy

E-mail: [email protected]

Giovanni ALOISIO3
PhD, University Professor, Department of Innovation Engineering,
Salento University, Lecce, Italy

E-mail: [email protected]

Abstract: Performance improvements in graphics hardware have made it possible to visualize

complex virtual environments and provided opportunities to interact with these in a more
realistic way. In this paper two different types of Virtual Reality applications for simulating a
billiards game are presented. In one application a commercial haptic interface is used to
provide a force feedback, thus rendering the interaction realistic and exciting to the user.
However, there are limitations due to the use of a commercial haptic device which has not
been specifically designed for this game and thus limits the workspace. Also, in the commercial
device, it is not possible to use the left hand when aiming and striking the ball, as you can in a
real game of billiards. In order to overcome these limitations another type of simulation has
been developed using a real billiard cue; its movements are reproduced in the virtual
environment using a visual marker detection system. No force feedback is provided to the
player.
In the game simulations the virtual environments have been built using the development
environment XVR in the first simulator and OpenSceneGraph in the second; rigid body
dynamics have been simulated utilizing the ODE and PhysX physics engines. ARToolkit was the
visual marker-based detection system utilized to replicate the movements of the real cue used
by the player in the virtual environment of the second simulator.

Key words: simulations; billiards game; virtual environments

440
 Software Analysis

In the field of computer entertainment new technologies have made it possible to

generate new forms of human-computer interaction where some bodily feedback is
provided, be it vibration or other, which is popular with players.
Haptic feedback in virtual environments makes it possible to increase the overall
realism of a simulation by improving the user’s experience and providing a deeper sense of
being in control of the game and of participation.
In this paper two different types of Virtual Reality simulations of the billiards game
are presented. The first uses a haptic device, in order to provide the user with an interactive
and realistic interaction. The force feedback is provided by means of a commercial haptic
interface and in this way it is possible to strike the billiard ball and to feel the contact
between cue and ball. 4
The second, in order to overcome the limitations due to the use of a commercial
haptic device which has not been specifically designed for the billiards game, uses a different
type of simulation which has been developed using a real billiard cue.
By means of a visual maker detection system the cue movements are replicated in
the virtual environment, but no force feedback is provided to the player.
Billiards game simulations have been developed both with and without the force
feedback sensation.
Gourishankar presents the HAPSTICK, a high fidelity haptic simulation of a billiards
game.5 The system incorporates a low cost interface designed and constructed for the haptic
simulation of the billiards game; the device allows motion in three degrees of freedom with
haptic feedback along the translation.
Takamura et al. present a billiards game simulation and the method used in this
research contributes to making the game extremely realistic.6
Visual markers are widely used in Augmented Reality (AR) applications. Currently
there are several different types of based marker tracking systems.
Zhang et al. compare several marker systems all using planar square coded visual
markers. They present the evaluation results, both qualitatively and quantitatively, in terms
of usability, efficiency, accuracy, and reliability.7
Wilczynski et al. describe internal structure and potential applications of a newly
constructed system for rapid game development in augmented environments. A description
of separating marker recognition and display engine of behaviour is provided.8
Ohshima, et al. present AR2 Hockey where two users wear see-through head
mounted displays to play an AR version of the classic game of air hockey and share a
physical game field, hockey sticks and a virtual puck to play in simultaneously shared
physical and virtual space.9

The Billiards Game Simulation Based on Force Feedback

In the first type of simulation of the billiards game, in order to make the game as
interactive and realistic as possible for the user, a force feedback is provided and it is
possible to strike the billiard ball and to feel the contact between cue and ball. By means of
a commercial haptic interface (PHANTOM Omni) a force feedback is provided, thus
rendering the interaction realistic and exciting to the user.
In the game simulation it is possible to distinguish three different types of
modelling: graphical, physical and haptic.

441
 Software Analysis

The graphical modelling consists of a set of 3D objects built using 3D Studio and
imported into the XVR development environment where they are managed using the XVR
scenegraph. An example of billiards with five skittles has been implemented.
Since in the real game it is possible to use your left hand when aiming and striking
the ball, in the play modality it is possible to fix the cue movement in the desired direction in
order to allow a more careful aim and a more stable interaction in the virtual environment.
In addition it is possible to choose the force amplification with which the ball is hit.
Each object of the scenegraph is modelled from the physical point of view defining
the geometry, the mass, the inertia, the stiffness and the contact friction with another one.
The ODE library is used to carry out the physical modelling definition and to define the
dynamics for simulating the billiards game.
Regarding the haptic modelling of the objects that are present in the virtual scene,
the utilization of the OpenHaptics library makes it possible to exercise control at a lower
level of the haptic interface. The cue is modelled as a rigid body and, in the play modality, its
position and orientation are linked, using a spring-damper system, to the position and
orientation of the haptic interface stylus.
Figure 1 shows the interactions with the virtual environment using a haptic
interface.

Figure 1. Haptic interaction

The limitations of the simulation are due to the use of a commercial haptic device
which has not been specifically designed for the billiards game. Because of the limited
workspace of the haptic device used, it is not possible to perform some shots, which, in the
real game, require wide movements in order to be carried out. In addition, it is not possible
to use your left hand in order to stabilize the cue and to obtain a more precise stroke, as
would happen in a real game of billiards. For this reason some modifications have been
introduced in the simulation; in particular it is possible to fix the chosen direction of the cue
during the strike and also to decide on the force amplification with which to hit the billiard
ball.

442
 Software Analysis

The Billiards Game Simulation Based on Marker Detection

In the second simulator of the billiards game, the player is not provided with a force
feedback because a real cue is used instead of a haptic interface.
By means of a marker detection system the movements of the real cue are
replicated onto the virtual one and this is able to interact with the other objects on the virtual
billiards table.
In this way, although players cannot feel the contact with the virtual ball, they can
carry out all the game procedures with a real cue and, as in the real game, they can use
their left hand in order to stabilize the cue and to obtain a more precise stroke.
Figure 2 shows a game phase using the developed billiards game simulator.

Figure 2. A billiards game phase

Regarding the construction of the virtual environment, the same models utilized in
the first simulator have been imported in OpenSceneGraph, the 3D graphics toolkit used in
this simulation.
OpenSceneGraph is an open source high performance and cross platform 3D
graphics toolkit written in Standard C++ and OpenGL; it is used in many flight simulators,
games and virtual reality visualization systems. It includes a wide range of features among
which there is a complete scene graph, support for a wide range of image and 3D model
formats.10
OpenSceneGraph is more compatible with ARToolkit, the software utilized to
manage the interactions in the virtual environment, and for this reason it has been chosen
over XVR.
To implement the dynamics of the rigid bodies that make up the virtual game
environment PhysX was the preferred choice.
The NVIDIA PhysX SDK is a physics engine used in a wide variety of console games
and game engines.11 Like ODE it allows rigid body dynamics simulation and collision
detection; in addition it offers a wide range of other features such as simulation of
deformable objects, advanced character control, articulated vehicle dynamics, cloth and
clothing authoring and playback, advanced fluid simulation.

443
 Software Analysis

PhysX is free for non-commercial and commercial use on PC platforms, but it is not
open source like ODE.
The visual marker-based detection system which was utilized in order to replicate
the movements of the real cue used by the player in the virtual environment is ARToolkit.
ARToolkit is a software library for building Augmented Reality applications and uses
square markers each carrying a unique pattern which is a planar bitmap enclosed by a black
border.12
Pattern recognition proceeds in two stages: recognition of the pattern boundaries
and correlation of the interior pattern with the patterns stored in a database.
These markers are observed by a single camera and the tracking software uses
computer vision techniques to calculate the marker position and orientation from the
captured image.
Markers can be used as a tangible interface to handle virtual artefacts or as user
interface elements. Tracking is impeded whenever the marker to be tracked is not fully and
clearly visible within the camera image; chances of full visibility can be improved by using
several markers fixed to a rigid object.
The offsets between the markers must be well-known and there must be some
components in the application which are able to calculate the final position of the object
from the valid tracking input.
The accuracy of tracking depends on many parameters in the processing chain: the
quality of the camera images, calibration of the camera, lighting, size and visibility of the
reference marker, the size of the marker to be tracked. If only one of these factors is not
optimally set, the results of tracking may be inaccurate or even unusable.
In the developed simulation of the billiards game it is possible, by means of a
webcam, to detect a marker grid used to define the position of the reference system with
respect which the movements of the real cue are calculated; these movements are detected
by means of a marker placed on the cue.
Figure 3 shows the interactions with the virtual environment using a marker-based
detection system.

Figure 3. The marker-based detection system

444
 Software Analysis

The use of a second marker on the cue was not considered because it would have
had to be placed in the visual field of the camera and thus close to the other one. This
solution is not feasible because the second marker would impede the cue movement during
the stroke.
The movements of the real cue are replicated on the virtual one that is modelled as
a physical body provided with mass of its own. The force applied to the ball is calculated by a
physic engine and is based on the speed of the real cue during the stroke.
The utilization of a physic engine as PhysX permits the modelling of the physical
proprieties of the virtual objects and hence defines their dynamic behaviour by means of
masses, frictions, etc.
Without the utilization of a haptic device, the force feedback due to the contact
between cue and other objects of the billiard table is lost, but the use of a real billiard cue
overcomes the limitations produced by the use of a commercial haptic interface which is not
specific to the billiards game.
A marker-based detection system was preferred to another type of tracking system,
such as an optical tracker, because it provides a solution which is both cheap and simple to
build.

Evaluation Test

This marker-based simulator, based on a marker detection system, allows the

player to handle a real billiard cue and thus to carry out all the strokes permitted in the real
game, but no force feedback is provided to the player.
In order to validate the simulator, some tests have been carried out in order to
check if the system is also able to detect the rapid strokes normally made in the real game.
To evaluate the performances of the tracking method based on marker detection, a
test application has been developed able to store the following positions of the billiard cue
detected by means of the tracking system during the stroke. This application makes it
possible to draw the trajectory obtained from the following positions detected by the tracking
system and to compare it with the linear path of the real cue during a stroke.
In this way it is possible to evaluate the ability of the system to detect the cue
positions in the cases of slow and fast strokes and to estimate the validity of the chosen
method. The data are also stored for future processing.
In the test phase ARToolKit Plus was chosen for use and just one marker to define
the position of the reference system with respect to the movements of the real cue; in this
way it was possible to achieve a higher degree of accuracy in marker detection and a
reduction in processing time. These improvements could easily be transferred to the
application.
The carried out tests highlight that the detection system is able to correctly register
the billiard cue trajectory in the case of slow strokes; however, when a rapid stroke occurs,
the number of detected cue positions decreases and the real trajectory departs slightly from
the ideal one. Figures 4 and 5 show the following positions of the tip cue detected by the
tracking system in the case of a slow stroke and the difference between the ideal (purple
line) and real (yellow line) trajectories.

445
 Software Analysis

Figure 4. Following positions of the tip cue in the case of a slow stroke

Figure 5. Difference between the virtual and real trajectories in the case of a slow stroke

Figures 6 and 7 show the following positions of the tip cue detected by the tracking
system in the case of a rapid stroke and the difference between the ideal (purple line) and
real (yellow line) trajectories.

Figure 6. Following positions of the tip cue in the case of a fast stroke

446
 Software Analysis

Figure 7. Difference between the virtual and real trajectories in the case of a fast stroke

Future Work

The analysis carried out to obtain a first validation of the marker-based billiards
game simulation is qualitative and the test results highlight that the method used to detect the
cue movements is not optimal.
It is probable that the use of a webcam provided with higher frame rate and
resolution and a more appropriate lighting of the game area would be useful in obtaining
better results.
An improvement in the simulation could be obtained by changing the modality of
the stroke and splitting it into two different phases; in the first phase of the stroke only the
movement of the billiard cue would be detected whereas in the second one the previously
acquired data would be processed in order to obtain the correct force to apply to the ball.
In this way it would be possible to detect and to correct the errors due to the tracking
system, but it is necessary to verify if the delay due to this processing remains enough short
during the simulation.
In addition, a quantitative analysis could be obtained by means of a comparison
with measurements obtained using a more accurate tracking system, such as an optical
tracker, where the margin of error is well known.

Bibliography

1. De Paolis, L., Pulimeno, M. and Aloisio, G. The Simulation of a Billiard Game Using a Haptic
Interface, in Roberts D. J. (ed.) “Proceedings of the 11th IEEE International Symposium
on Distributed Simulation and Real Time Applications”, Los Alamitos, CA: IEEE Computer
Society, 2007, pp. 64-67
2. Gourishankar, V., Srimathveeravalli, G. and Kesavadas, T. HapStick: A High Fidelity Haptic
Simulation for Billiards, in “Proceedings of the Second Joint EuroHaptics Conference
and Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems,
World Haptics 2007”, Los Alamitos, CA: IEEE Computer Society, 2007, pp. 494-500
3. Ohshima, T. et al. AR2Hockey: A Case Study of Collaborative Augmented Reality, in
“Proceedings of the 1998 Virtual Reality Annual International Symposium”, Los Alamitos,
CA: IEEE Computer Society, 1998, pp. 268-275
4. Takamura, Y. et al. A Virtual Billiard Game with Visual, Auditory and Haptic Sensation, in
Pan, Z. (ed.) “Technologies for E-learning and Digital Entertainment”, Heidelberg,
Springer Berlin, 2006, pp. 700-705

447
 Software Analysis

5. Wilczyński, Ł. and Marasek, K. System for Creating Games in Augmented Environments, in

“Proceedings of the 2007 International Conference on Multimedia and Ubiquitous
Engineering”, Los Alamitos, CA: IEEE Computer Society, 2007, pp. 926-931
6. Zhang, X. Fronz, S. and Navab, N. Visual Marker Detection and Decoding in AR Systems: A
Comparative Study, in “Proceedings of the 2002 International Symposium on Mixed
and Augmented Reality”, Los Alamitos, CA: IEEE Computer Society, 2002, pp. 97-106
7. * * * ARToolKit Home Page, http://www.hitl.washington.edu/artoolkit/
8. * * * Nvidia PhysX Main Page, http://www.nvidia.com/object/nvidia_physx.html
9. * * * OpenSceneGraph Home Page, http://www.openscenegraph.org/projects/osg

1
Lucio Tommaso De Paolis is an Assistant Professor of Information Processing Systems at the Department of
Innovation Engineering of the University of Salento (Italy).
He received a Degree in Electronic Engineering from the University of Pisa (Italy) and since 1994, first at the Scuola
Superiore S.Anna of Pisa and then at the University of Salento, his research interest has concerned the study of the
interactions in the virtual environments and the development of Virtual Reality and Augmented Reality applications
in medicine and surgery.
He is a member of the Society for Medical Innovation and Technology (SMIT), member of the Southern Partnership
for Advanced Computational Infrastructure Consortium (SPACI Consortium) and member of the Italian Movement of
Modelling and Simulation (MIMOS).
He teaches Computer Science at the Faculty of Sciences of the Salento University.

2
Marco Pulimeno is contract worker at the Salento University where he obtained a Bachelor Degree in Computer
Science Engineering. His main interests include Augmented and Virtual Reality technologies applied to gaming and
medicine.

3
Giovanni Aloisio is Full Professor of Information Processing Systems at the Engineering Faculty of the University of
Lecce. His research interests are in the area of High Performance Computing, Distributed and Grid Computing and
are carried out at the Department of Innovation Engineering of the University of Lecce. He has been a co-founder of
the European Grid Forum (Egrid) which then merged into the Global Grid Forum (GGF). He has founded SPACI
(Southern Partnership for Advanced Computational Infrastructures), a consortium on ICT and grid computing among
the University of Lecce, the University of Calabria and HP Italia. The Consortium is a follow-up of the SPACI project
funded by the Italian Ministry of Education, University and Technological Research, to pursue excellence in the field
of Computational Science and Engineering. He is the author of more than 100 papers in refereed journals on
parallel & grid computing.

4
De Paolis, L., Pulimeno, M. and Aloisio, G. The Simulation of a Billiard Game Using a Haptic Interface, in
Roberts D. J. (ed.) “Proceedings of the 11th IEEE International Symposium on Distributed Simulation and Real Time
Applications”, Los Alamitos, CA: IEEE Computer Society, 2007, pp. 64-67

5
Gourishankar, V., Srimathveeravalli, G. and Kesavadas, T. HapStick: A High Fidelity Haptic Simulation for
Billiards, in “Proceedings of the Second Joint EuroHaptics Conference and Symposium on Haptic Interfaces for
Virtual Environment and Teleoperator Systems, World Haptics 2007”, Los Alamitos, CA: IEEE Computer Society,
2007, pp. 494-500

6
Takamura, Y. et al. A Virtual Billiard Game with Visual, Auditory and Haptic Sensation, in Pan, Z. (ed.)
“Technologies for E-learning and Digital Entertainment”, Heidelberg, Springer Berlin, 2006, pp. 700-705

7
Zhang, X. Fronz, S. and Navab, N. Visual Marker Detection and Decoding in AR Systems: A Comparative
Study, in “Proceedings of the 2002 International Symposium on Mixed and Augmented Reality”, Los Alamitos, CA:
IEEE Computer Society, 2002, pp. 97-106

8
Wilczyński, Ł. and Marasek, K. System for Creating Games in Augmented Environments, in “Proceedings of
the 2007 International Conference on Multimedia and Ubiquitous Engineering”, Los Alamitos, CA: IEEE Computer
Society, 2007, pp. 926-931

9
Ohshima, T. et al. AR2Hockey: A Case Study of Collaborative Augmented Reality, in “Proceedings of the 1998
Virtual Reality Annual International Symposium”, Los Alamitos, CA: IEEE Computer Society, 1998, pp. 268-275

10
* * * OpenSceneGraph Home Page, http://www.openscenegraph.org/projects/osg

11
* * * Nvidia PhysX Main Page, http://www.nvidia.com/object/nvidia_physx.html

12
* * * ARToolKit Home Page, http://www.hitl.washington.edu/artoolkit/

448
 Software Analysis

DATABASE SECURITY - ATTACKS AND CONTROL METHODS

Emil BURTESCU1
PhD, Associate Professor, Department of Accounting and Management Informatics,
University of Pitesti, Pitesti, Romania

E-mail: [email protected], [email protected]

Abstract: Ensuring the security of databases is a complex issue for companies. The more
complex the databases are the more complex the security measures that are to be applied are.
Network and Internet connections to databases may complicate things even further. Also, each
and every additional internal user that would be added to user base can create further serious
security problems. This pupose of this paper is to highlight and identify the main methods and
facets of attack on a database, as well as ways to deflect attacks, through focusing on the
delicate issue of data inference.

Key words: attack; control; impact; inference; security

Like all tangible assets that have to be protected by a company, valuable

information stored in its computer system database is probably the most precious of assets
of the company that must be protected.
Safety measures must be an integral part of any database, right from the start, at
the inception and design phase. Modern approaches employed to assure the security of
databases address security and protection defenses at all levels: physical, network, host,
applications and data.
It goes without saying that the first of such measures has to be applied starting at
the physical level and to then progress right through, reaching the data level at the other
end. Initially, companies have had a rather simplistic approach, mainly due to primitive and
rudimentary nature of earlier attacks, as well as the simple nature and construction of the
then prevalent networks with very limited complexity if any, and did therefore focus on
assuring security at the physical level. That then involved basic measures such as limiting
access to locations that only authorized personnel may have access to data.
More recently, due to the rapidly changing and increased size as well as
complexity and expansion of company information systems, AAA type measures began to be
used (Authentication, Autorisation, Access).
Currently the necesary security measures are far more complex. These are meant to
stop the highly sophisticated attacks from external attackers, and especially, from those
who may very well have access to the company’s internal network.

449
 Software Analysis

1. Classical attacks

The focus of attacks on the company’s databases are motivated by the following
factors:
• Databases are the mass of information which the company works with;
• Databases can reveal private data by processing public data.
Database security is relative in the next situations:
• Theft and fraud;
• Loss of confidentiality/privacy;
• Loss of privacy;
• Loss of integrity;
• Loss of availability.
The hazards which make these things happen are due in large amount to
deliberate human action. Natural type hazards or random events have an impact only on
data integrity and availability.
To ensure a minimum security of the databases the following requirements must be
satisfied:
• Physical integrity of databases;
• Logical integrity of databases;
• The integrity of each element which composes the database;
• Access control;
• User identification;
• Availability.
The physical and logical integrity of databases will require the focus of efforts for
protecting the physical integrity of databases, especially the recordings against destruction.
The easiest way to do that is represented by regular backups.
The integrity of each element forming the database will assume that the value of each
field may be written or changed only by authorized users and only if there are correct values.
The access control is being done taking into consideration the restrictions of the database
administrator. DBMS will apply the security policy of the database administrator (DBA).
This must meet the following requirements:
• Server security. Server security involves limiting access to data stored on the server.
It is the most important option that has to be taken in consideration and planned
carefully.
• Connections to the database. Using the ODBC will have to be followed by
checking that each connection corresponds to a single user who has access to data.
• Access control table. The access control table is the most common form of
securinga database. An appropriate use of the table access control involves a close
collaboration between the administrator and the base developer.
• Restriction tables. Restriction tables will include lists of unsure subjects who could
open set off sessios.
Secure IP addresses. Some servers may be configured to receive only queries from hosts
that are in a list. Oracle servers allow blocking queries that are not related to the database.
Cancellation of the Server Account. The ability to suspend an account when guessing the
password is tried after a predefined number of attempts (usually 3).

450
 Software Analysis

Special tools. Special programs such as Real Secure by ISS which will alert in case of
intrusion attempts. Oracle has an additional set of authentication methods: Kerberos
security; Virtual private databases; Role-based security; Grant-execute security;
Authentication servers; Port access security.
User identification will allow at any time to be known who does anything in the system. All
the operations performed by users will be stored and will form a history of access. Checking
the history of all hits is sometimes hard and requires a considerable workload.
Availability will allow the required data to be available for an authorized user.

2. Attacks specific to the databases

Unlike other types of data, databases may be subject to unilateral actions, in which
an unclassified user has access legitimately to public information but on which he is able to
infer classified information. These types of actions are called inference.
After such actions two situations are distinguished which lead to the disclosure of
secret data from public data: data aggregation and association.
Two cases of inference which often appear in databases: data aggregation and
data association.
Data aggregation problem arises whenever a set of information is classified at a higher
level than individual levels of involved data.
Example: Military field - Individual location of vessels is unclassified, but the overall
information about the location of the entirefleet is secret. Commercial - Total sales reports of
different branches of the company can be seen as less confidential than the global reports of
the company.
Data association problem arises whenever two values taken together are classified at a
higher level than the one of each value.
Example: The list containing the names of all employees and a list containing salaries are
unclassified, and a combined list containing the names and the salaries of employees is
considered classified.
A first step in countering these types of attacks is the protection of sensitive data-
data that must not be made public. It is considered as being sensitive data facts that are
inherently sensitive, from a sensitive source, are declared sensitive, come from a recording
or an attribute which is sensitive or are not sensitive in relation with other sensitive data.
Applying one or more methods of attack, and in combination with a weak
protection of databases, several sensitive data types may be displayed:
Accurate data. When the database does not implement any protection mechanism, the
extracted data is exatcly the exepcted ones. Queries are simple and obvious.
Bound data. In this situation an attacker can determine the range of values which the
searched value can have.
Existing data. Data are classified but which can be emphasized that the existence by a
process of inserting data protection mechanisms, operation refused by the protection
mechanisms of the database because the data already exist.
Negative data. After some seemingly innocent queries sensitive data can be displayed. A
query will be able to display data whose existence is not known, these being sensitive.
Probable data. Their existence is highlighted by complex attacks.
The success of attacks on databases relies heavily on the skills and training of the
attacker and less on the automation mechanisms of attack. They use pretty much their

451
 Software Analysis

knowledge and statistical tools, and because in this these attacks are also called statistical
attacks or statistical inference attacks.
An attacker, after he passed all levels of protection and reached the database, he
will try progresively a series of attacks: direct, indirect and by tracking.
Direct attacks are obvious attacks and are successful only if the database does not
implement any protection mechanism. The displayed results will be the ones required and
expected. If this attack fails then the attacker moves to the next.
Indirect attacks are attacks that are executed when it is desired the extraction of other data
than tose that are displyed. Combinations of queries are used some of them having the
purpose to cheat thesecurity mechanisms.
The tracking attack is applied to the databases that have impemented a supression
mechanism for the claims that have dominant results. This type of attakc is used against
databases that have short answers to queries. Attacks are based on the principle to which if
a direct query has as result a small number of answers, the denial of the main claim will
result in zero. If the answer to a complex claim is not displayed due to the supression
mechanism for claims with dominant result, then the database will be queried with a set of
claims and the answer to these claims will be studied, following that from these sensitive
data to be extracted. In literature, this type of attack is called Linear System Vulnerability.

3. The risk

Focusing efforts to ensure database security must be done considering firstly the
impact the loss of data has on the business. The final purpose must bear in mind the
assurance of confidentiality, integrity, availability and data non repudiation. If the first three
objectives are already classic, the last one, the non repudiation, it is necesarry in electronic
transactions for confirming authenticity.
A quantitative approach of the risk is preferable than a qualitative approach
because it offers a more tangible value of the situation. Even so we will still work with
subjective data, estimated after an evaluation process.
If in the case of hardware loss it will be easier to estimate the loss using the cost of
replacing the component, in case of a data loss the operation is far more complex. In this
case we will discuss about costs for recovery. For a quantitative approach we will start from
the formula for calculating the risk:

Risk = Impact x Probability

To estimate the impact we have to ask ourselves if: data can be rebuilt or restored;
how long does it take to rebuild data; it is because of a deliberate action or because of
accidental causes; the lsot data have special character (military, secret service, confidential).
n
Impact a = ∑ Impact i where: Impacta – Total impact for asset a; i – impact zone, (1 to 4,
i =1

confidentiality, integrity, avalability and nonrepudiation).

Calculus of impact value is exemplified in next table.

452
 Software Analysis

Table 1. Impact value calculus

Impact zone Impact value Observations
(USD)
Confidentiality 5 000 Loss due to the data theft. Very difficult evaluation. It implies the
cooperation of several departments.
Integrity 1 000 Loss due to data spoilage. It involves costs for checking data
integrity. Complex operations.
Avalability 800 Loss due to the unavailability of data. It involves costs for
restoration and availability of data. Complex operations and under
pressure.
Nonrepudiation 100 Loss due to orders denial.
Total 6 900

The probability to happen an incident on the databases must be estimated by the

analysis team for the security risk. The probability of generating natural events that can
disrupt the smooth functioning of databases may be provided by the organizations in the
field. Another category is represented by the threats specific to every company and which are
linked to the human factor. Once the risk management is mature the estimation of
probability of some events to occur would be more precise. Creating diagrams of evolution
of the risk would help the company to concentrate its efforts one the areas that are the most
affected and enhance methods of control on these locations.

4. Control methods

The classical methods of ensuring database security, the partition of database,

cards, encryption, etc. Are able mostly to accomplish their tasks. Yet these are not sufficient.
Supposing we have already implemented the security mechanisms which permit us
to know who the user is, then the only thing left is that we have to see what the user does.
Using appropriate mechanisms for logging and auditig operations we will be able to see
what every user has done and in case of incident the user to be held responsible.
Once this is made, we can go to the phase of attack control. This will involve the
implementation of a mechanism that will not permit displaying sensitive data.
The options that can be chosen for such mechanism are the following:
Suppressing the applications with sensitive results. The requests for access for
database elements that have as result displaying sensitive results are rejected without any
response.
Results approximation. The results of request will be approximated in such way that the
attacker will not be able to determine the exact values. In the case of such request the
system will be able to display results close to the real ones.
Limiting the results of a request that reveals sensitive data. Limiting the result of a
request which reveals sensitive data can be done in the case in which this is 1 (one).
Combining results. Combining the results from several request will create even a greater
confusion for the attacker.
All these can be embedded in a monitor type mechanism which will implement de
security policy of the company. Access to data will be done according to the user’s
classification and data classification. There is sometimes confusion between the user and the
end-user. An end-user must have access only to one or more programs that run

453
 Software Analysis

applications. A user is defined as being that person who has access to a computer system.
The end-user is actually an operator and so it should stay.
Persons who work in security field or are on the other side of the barricade agree
that a security ensurance system must resist 3-5 days to fulfill its purpose.
Other security controls to ensure the security of databases include control elements
that are not based on the computer. Here we include policies, agreements and other
administrative control elements different than the ones who sustain control elements
based on the computer. From this category we have:
• Security policy and emergency situations plan;
• Staff control;
• Placing the equipment in safe conditions;
• Escrow agreements;
• Mainentance agreements;
• The physical control of access.

5. Conclusions

Database security presents features that must be seriously taken into account.
The first option, for a secure database is represented by its optimal protection.
Ensuring database security must be done from outside ton inside, this involving
ensuring security starting from the physical level and ending with the data level (physical,
network, host, applications and data).
Databases are a favourite target for attackers because of the data these are
containing and also because of their volume. Datawarehouse is the ultimate goal.
Efforts to ensure database security are considerably higher than for the other types
of data. It is easier to implement an access list for a great number of files than an access list
for the elements of a database.
Database security mechanisms should not irritate their users.

References

1. Burtescu, E. Problems of Inference in Databases in „Education, Research & Business

Technologies“. The Proceeding of the 9th International Conference on Economic
Informatics, INFOREC Printing House, Bucharest, 2009
2. Burtescu, E. Databse security, in „Knowledge Management. Projects, Systems and Technologies“,
International Conference „Knowledge Management. Projects, Systems and
Technologies“, Bucharest, November 9-10, 2006, INFOREC Printing House, Bucharest,
2006
3. Hsiao, S.B. and Stemp, R. Computer Security, course, CS 4601, Monterey, California, 1995
4. McCarthy, L. IT Security: Risking the Corporation, Prentice Hall PTR, 2003
5. Proctor, P.E. and Byrnes, F.C. The Secured Enterprise, Prentice Hall PTR, 2002
6. * * * Security Complete, Second Edition, SYBEX Inc., 2002

1
Emil BURTESCU has graduated the Polytechnics University of Bucharest, Faculty of Aerospace Engineering. He
holds a PhD diploma in Economic Cybernetics and Statistics at Faculty of Cybernetics, Statistics and Economic
Informatics, Bucharest Academy of Economic Studies.

454
 Quantitative Methods Inquires

A MATHEMATICAL MODEL FOR TUMOR VOLUME EVALUATION

USING TWO-DIMENSIONS1

John P. FELDMAN
MSc, Department of Nuclear Engineering, Ben-Gurion,
University of the Negev, Beer-Sheva, Israel

E-mail:
Ron GOLDWASSER
MSc, Department of Nuclear Engineering, Ben-Gurion,
University of the Negev, Beer-Sheva, Israel

E-mail: [email protected]
Shlomo MARK2
PhD, Senior Lecturer, Negev Monte Carlo Research Center (NMCRC) and
Department of Software Engineering, Shamoon College of Engineering, Beer Sheva, Israel

E-mail: [email protected]
Jeremy SCHWARTZ3
Research Assistant, Negev Monte Carlo Research Center (NMCRC) and
Department of Software Engineering, Shamoon College of Engineering, Beer Sheva, Israel

E-mail: [email protected]
Itzhak ORION4
PhD, Senior Lecturer, Department of Nuclear Engineering, Ben-Gurion,
University of the Negev, Beer-Sheva, Israel

E-mail: [email protected]

Abstract: Many recent papers present different ways to show the volume of a tumor from a
few two-dimensional images. The three-dimensional fundamental shape of tumors is
assumed to be a hemi-ellipsoid as presented in different studies. The three measurements
were essential for tumor volume calculations: length, width, and height. Tumor volume
measurements task is a very intensive routine is cancer research. Recent papers present how
to reconstruct the 3-D tumor from a set of 2-D images, this in order to find the tumor volume.
In this paper we report on a new approach to calculating the volumes based on
measurements of two dimensions, length and width, after having identified a statistical
constant that replaced the need of measuring the tumor height. Results: Our new method was
examined on a subcutaneously implanted tumor placed on a mouse's thigh. The width,
length, and height of tumors were measured, in four groups of BALB/c mice, using a digital
caliper. The tumor dimensions were periodically measured for several weeks. It was shown
that this new model can assist in tumor volume measurements using digital images, and in CT
scan tumor size assessments.
Key words: tumor volume; tumor reconstruction; tumor imaging; hemi-ellipsoid; mice model

455
 Quantitative Methods Inquires

1. Introduction

When using medical imaging that involve radiation such as computed tomography,
it is important to minimize the patient exposure. The high exposure is delivered due to the
need of a number of images to reconstruct the tumor shape, and dimensions (Junzhou,
2006).
The three-dimensional fundamental shape of tumors is assumed to be a hemi-
ellipsoid in different studies. Researches that use the assumption of a hemi-ellipsoid tumor
shape were published regarding breast cancer (Wapnir, 2001), prostate cancer (Sosna,
2003) (Egorov, 2006) cervical cancer (Mayr, 2002), glioma cancer (Schmidt, 2004), and
others (James, 1999).
A typical ellipsoid volume is giving by:
π
V= (length ) ⋅ ( width ) ⋅ (height ) (1)
6
This study aims to assist in tumor volume measurements by developing a new
model that reduces the essential number of dimensions for the volume, and therefore
reduces the number of images needed.
The new mathematical model for tumor volume measurement was investigated
using a mice model, which is described in the methods section.

2. Methods and Materials

The mice model

In order to examine a new method of tumor volume measurements a
subcutaneously implanted tumor was placed on a mouse's thigh. The width, length, and
height of tumors were measured, in four groups of BALB/c mice, using a digital caliper. In
cancer treatments, determination of the growth rate of tumor volume as a function of time is
a standard method of determining the efficiency of a particular treatment.
Since changes in the growth rate reflect the efficiency or inefficiency of a treatment,
extreme precision in the measurements is critical. The KHJJ tumor line was derived from a
primary mammary tumor arising in a BALB/c mouse, after implantation of a hyperplastic
alveolar nodule (Rockwell, 1972). Four groups of mice—26 individuals altogether—were
tested in the experiments. All the mice were of the same BALB/c type and of similar size (28
± 1.4 SD gram average weight). There were two separate groups for each gender (see Table
1). The mice received a tumor transplant on the thigh. The mice were kept and treated
according to the Ben-Gurion University of the Negev guidelines for treating animals in
scientific experiments.

Table 1: Members of each group of mice and the period of the measurements
Group Research period Number of mice in group Gender
1 63 days 8 male
2 39 days 2 male
3 36 days 5 female
4 43 days 11 female

456
 Quantitative Methods Inquires

Preparation of the tumor segments

The KHJJ tumor segments were prepared from a KHJJ tumor about 200 mm3 in
size. This tumor was taken from a mouse that had been sacrificed moments before
dissection. The tumor had to be without signs of necrosis and with smooth margins. The
tumor was separated properly from any remaining healthy tissue and washed three times
with PBS. Then the tumor was cut through the middle to ensure that there was no necrotic
tissue inside. The segments were prepared by cutting the tumor into 1 mm3 sections for
transplantation. These segments were kept moist by dripping PBS on them until implantation.
This procedure was carried out in sterile conditions under a sterile hood.

Tumor transplantation
Each mouse, before the transplant, was anesthetized with a low dose of 80 ml/kg
Ketamine and 8 ml/kg Xylazine anesthetics for about 40 minutes. The tumor segments were
prepared for dissection a few moments before the process.
The mouse’s right leg was shaved, the thigh skin was pulled up with forceps, and
then a subcutaneous slit was made along the skin. A 1 mm3 tumor segment was inserted
using a trocar into the small pocket under the skin on the thigh. A small, flat stick covered
with antibiotic ointment was pressed on the skin slit while the trocar was removed.

Tumor volume measurements

Once the tumor became palpable 5–10 days after transplant, its size was measured
using a digital caliper. Only one person measured all the tumors in the experiments to
prevent observation differences, since it was found that measurements by more than one
person can lead to different results. The tumor was measured every other day, within four
hours before or after the previous measurement. The tumor was measured between the skin
surface layers. The length and width were measured with an accuracy of 0.01 mm using a
digital caliber. The length was measured along the imaginary longitude of the leg; the width
was measured in the direction of the latitude. The height was measured between the leg
surface layer and the upper skin of the tumor. The caliper was placed perpendicular to the
tumor so that the height could be measured properly. The tumor was measured from a
volume of about 50 mm3 until it had grown up to 1500 mm3. Each volume measurement
was repeated three times for verification.
Measuring tumor volume is problematic mainly because of inaccuracy in the
measurement of tumor height, which contributes the largest error to volume results. The
difficult part in achieving precise measurements is determining where to position the caliper
in order to measure height. It is clear that a new approach eliminating the need to measure
the height, while still providing a precise assessment of the tumor volume, could be very
helpful. Tumor length and width can be measured very accurately because they can be
observed directly. The 0.01 mm instrumental error associated with the caliper can lead to a
0.5%–1.0% error in a volume of 50 mm3. The intensive repetitions in this study indicate that
the total error for length and width is about 0.1 mm each, leading to a 3% error in the
resulting volume. The tumor height error is very large when measuring with a caliper,
around 0.5 mm, and can add a 7% error to the volume value. The total estimated volume
error is around 10% (Dethlefsen 1968).
In order to minimize the volume error, we suggest an approach that relies only on
length and width measurements to calculate volume. Such a calculation may a priori reduce

457
 Quantitative Methods Inquires

the uncertainty regarding the volume if a geometrical dependence of volume on length and
width can be established:

Volume = f ( width, length ) (2)

3. Results and Discussion

The dimensions of the tumor in the mouse model were measured only from the
time it reached a volume of 50 mm3 until it had grown to 1500 mm3.
Different tumor growth rates were observed for different mice. Figure 1 shows
examples from two mice groups—a male group and a female group. The graphs
demonstrate a change in tumor dimensions over time.

600
500
Female # 560
400
300
200
100
0
600
500 Female # 741
400
(a)
2

300
mm

200
100

600
500
Female # 752
400
300
Length xWidth
200 2
100 3Height
0
0 5 10 15 20 25 30
Days

600
500 Male # 359
400
300
200
100
0
600
500 Male # 416
400
(b)
2

300
mm

200
100
600
500
Male # 2539
400
300 Length xWidth
200 2
3Height
100
0
0 5 10 15 20 25 30 35 40 45 50 55
Days
Figure 1. (Length × width) and (3 × height2) over time. The red points
represent data pertaining to volumes that were found to be
irrelevant: (a) three female mice; (b) three male mice

458
 Quantitative Methods Inquires

We compared the average product of length and width with the square of height. A
rough fit was discovered between the two values, with the square of the height multiplied by
a factor of 3 being approximately equal to the length multiplied by the width:
3H 2 ≅ LW (3)
The examples in Figure 2 agree with Equation 3 for most of the points that
represent different volumes. In addition, it should be noted that the different tumor growth
rates did not affect the fit. A linear fit of all the results, shown in Figure 2, resulted in the
following equation:
H = 1.63 LW (4)
The correlation represented by Equation 4 was found to be high, with a linear
correlation coefficient of R = 0.919. For a normal distributed data set, the likelihood
estimator can be obtained by a least squares analysis (Alfassi, 2005). Therefore this
correlation coefficient shows best the validity of Equation 4.
20
All Mice

15
(mm)
1/2
(Length x Width)

10

5 Fit Y = C * X
C = 1.63 ± 0.01
R = 0.91888

0
0 2 4 6 8 10 12 14
Height (mm)

Figure 2. The linear fit of the square root of length × width corresponded to the
height in all the measurements of mice.

In order to determine accurately the relationship between H2 and LW, these values
were analyzed separately for each gender, with the following results:

H
Females: f = = 1.58 ± 0.01 (5)
LW
H
Males: f = = 1.69 ± 0.03
LW
A new formula for calculating tumor volume without the use of tumor height was
obtained from the analysis of the measurements:
π 3
V= f ( length ⋅ width ) 2 (6)
6

459
 Quantitative Methods Inquires

The new formula is based on some symmetry assumptions that inherent in the
classical volume formula, as the classical volume formula is a simple expression of an
ellipsoid volume. A comparison of the new volume calculation with the classical calculation
based on three dimensions, seen in Figure 3, shows no apparent difference in volume
values. The total mouse mass growth-rate usually differs between males and females. The
difference in the tumor growth-rate can be explained by the distinction in tumor growth-rate
between genders.
The total volume error can be reduced using this new method in tumors placed
subcutaneously to around 4%, compared to the 10% error that was obtained in the old-
fashioned volume measurements. The error bars in Figure 3 are larger for the classical
volume values and smaller for the new volume results.

1600
1600
Female
Male

1200

Classic Volume (mm )

1200

3
Classic Volume (mm )
3

800 800

400 400

(a) (b)

0 0
0 400 800 1200 1600 0 400 800 1200 1600
3 3
New Volume (mm ) New Volume (mm )

Figure 3. A comparison of the volume calculated according to the classical three-

dimensional formula with the volume calculated according to the new method.
The fit shows a good match between the results of the two methods: (a) in female
mice; (b) in male mice.

4. Conclusions

The new method for tumor volume calculations was studied using the mice model
described above. This model showed improved error estimation for the tumor volume. The
old method results were plotted against the new method results, shown in Figure 3. The
linear fit line for this graph obtained a slope of 1.00, proving the consistency of the new
method. The new method of calculating volume should reduce the error in the volume
measurements because it depends on visual measurements that can be accurately obtained.
Since the nominal errors for length and width are similar, using the new method, presented
in Equation 4, should limit the height error to about 1.5 times the error in length. This new
method, not only reduces the error of the tumor volume, but also reduces the number of
parameters needed to be measured down to two (length and width) from three (including
the height). The new method can be helpful in several cases where a digital photo of a
tumor can be taken, and may shorten the time needed for handling mice in the lab (see
Appendix).

460
 Quantitative Methods Inquires

This work presents a test-case study of the new mathematical method offered for a
tumor placed subcutaneously. These findings should encourage future studies towards
reducing the amount of medical imaging scans needed for internal tumor volume
reconstruction.

References

1. Alfassi, Z.B., Boger, Z., and Ronen, Y., Statistical treatment of analytical data, Boca Raton (FL),
CRC Press, 2005
2. Dethlefsen, L.A., Prewitt, J.M.S. and Mendelsohn, M.L., Analysis of tumor growth curves, J Natl
Cancer Inst., 40, 1968, pp. 389-405
3. Egorov, V., Ayrapetyan, S. and Sarvazyan A.P. Prostate Mechanical Imaging: 3-D Image
Composition and Feature Calculations, IEEE Transactions on Medical Imaging, Vol.
25, No. 10, October 2006
4. James. K., Eisenhauer, E., Christian, M., Terenziani, M., Vena, D., Muldal, A. and Therasse, P.
Measuring Response in Solid Tumors: Unidimensional Versus Bidimensional
Measurement, J Natl Cancer Inst, Vol. 91, No. 6, March 17, 1999
5. Junzhou, H., Xiaolei, H., Metaxas, D. and Banerjee, D. 3D tumor shape reconstruction from 2D
bioluminescence images, 3rd IEEE International Symposium on Biomedical Imaging:
From Nano to Macro. 6-9 April 2006, Biomedical Imaging, Macro to Nano, 2006
6. Mayr, N.A., Taoka, T., Yuh, W.T.C., Denning, L.M., Zhen, W.K., Paulino, A.C. et. al. Method and
timing of tumor volume measurement for outcome prediction in cervical cancer
using magnetic resonance imaging, Int. J. Radiation Oncology Biol. Phys., Vol. 52,
No. 1, 2002, pp. 14–22
7. Rockwell, S.C., Kallman, R.F. and Fajardo, L.F., Characteristics of a serially transplanted
mouse mammary tumor and its tissue-culture-adapted derivative, J Natl Cancer
Inst, 49, 1972, pp. 735-49
8. Schmidt, K.F., Ziu, M., Schmidt, N.O., Vaghasia1, P., Cargioli, T.G., Doshi, S. et. al. Volume
reconstruction techniques improve the correlation between histological and in
vivo tumor volume measurements in mouse models of human gliomas, Journal
of Neuro-Oncology, 68, 2004, pp. 207–215
9. Sosna, J., Rofsky, N.M., Gaston, S.M., DeWolf, W.C. and Lenkinski, R.E. Determinations of
prostate volume at 3-tesla using an external phased array coil Comparison to
pathologic specimens, Academic Radiology Vol. 10, Issue 8 , 2003, pp. 846-853
10. Wapnir, I.L., Barnard, N., Wartenberg, D. and Greco, R.S. The inverse relationship between
microvessel counts and tumor volume in breast cancer, Breast J., 7(3), May-Jun,
2001, pp. 184-8

Appendix
We developed a Windows-utility-program to insert a digital photograph of a tumor
(shown in Figure 4). This program could be used to measure tumor length and width, and
the tumor volume could be calculated using the new method presented in this paper. It
would also be possible to use the classical method of calculating volume, if desired, once the
tumor height has been determined. The caliper is not necessary to measure the tumor
volume if one uses this program with digital photographs of the tumor. The program can be
downloaded from http://www.bgu.ac.il/~iorion/.

461
 Quantitative Methods Inquires

Figure 4. An example of tumor volume measurements using the Windows utility program

1
Acknowledgments: We thank Prof. Brenda Laster of the Department of Nuclear Engineering, Ben-Gurion
University of the Negev, for allowing us to use her research labs.

2
Shlomo Mark is a senior lecturer at SCE Shamoon college of engineering. He earned his Ph.D. in nuclear engineering
and an M.Sc. in Biomedical Engineering and in Managing and Safety Engineering . He works in the Department
of Software Engineering. and he is the Head of the NMCRC - Negev Monte Carlo Research Center, Shamoon College of
Engineering. His main research interests are Scientific programming, Computational modeling for Physical,
environmental and medical applications, Monte Carlo Simulations, Develop, upgrade, and improved Monte Carlo based
codes, by using software engineering mythologies.

3
Jeremy Schwartz is a software developer with a broad range of programming experience from research,
consulting, and academic settings. He graduated in 2006 from North Carolina State University (USA) with
undergraduate degrees in Chemical Engineering and Computer Science. Mr. Schwartz specializes in writing highly
functional, well-documented and efficient code for both online and offline environments. Having written training
materials, developed data analysis tools, and rebuilt several websites from the ground up, he understands the
iterative process by which user requirements are transformed into deliverables.

4
Itzhak Orion is a senior lecturer at Ben-Gurion University. He earned a Ph.D. in nuclear engineering and an M.Sc.
in nuclear physics. He works in the Department of Nuclear Engineering. He visited several institutes around the
world as a collaborating scientist. He serve as an advisor to the Israeli atomic energy research centers in the
subject of radiation simulations. His main research interests are radiation, Monte Carlo simulations, computational
modeling and applications for medical physics.

462
 Quantitative Methods Inquires

A FORECASTING MODEL WITH CONSISTENT

ADJUSTMENTS FOR ANTICIPATED FUTURE VARIATIONS

Chin-Lien WANG1
PhD Candidate, Department of Industrial Engineering and Enterprise information,
Tunghai University, Department of Business Administration, Ling Tung University,
Taichung, Taiwan

E-mail: [email protected]

Li-Chih WANG2
PhD, University Professor, Department of Industrial Engineering and Enterprise information,
Tunghai University, Taichung, Taiwan

E-mail: [email protected]

Abstract: Due to the limitation of most statistical forecasting models ignoring contextual
information, judgmental adjustment is a widespread practice in business. However, judgmental
adjustment still suffers with many kinds of biases and inconsistency inherent in subjective
judgment. Our approach uses an adjustment mechanism concerning only with critical cue
factors evaluated with genetic algorithm to alleviate problems caused by collinearity and
insignificant sporadic variables usually arising in least square type estimators, and to derive
more realistic parameter estimation. In case there are anticipated variations in the forecasting
horizon and can’t be handled by the model alone, this adjusting mechanism, formulated in a
set of equations, can be used to assess mixed effect of cue factors consistently and effectively
without subjective judgment involved. Empirical results reveal that this adjustment mechanism
could significantly reduce MAPE of forecasts across seasons with improvement mainly coming
from large size adjustments.

Key words: judgmental adjustment; seasonal index realignment; genetic algorithm;

calendar effect

1. Introduction

Owing to the limitations of statistical forecasting methods generating forecasts

solely based on historical data [1-3]3, or not including critical explanatory variables, as
pointed out in [4], judgmental adjustments taking advantage of contextual information or
non time series information [5] become a widespread practice in business to improve
forecasting accuracy [5, 6-8], especially for model-based forecast in variable environments.
As [9] put it, the benefits should be greatest where series are subject to high noise and/or
where the signal is relatively complex. Blattberg and Hoch [10] also argue that forecasters

463
 Quantitative Methods Inquires

can use econometric models effectively only if they have a built-in adjustment mechanism to
capture the changing environment. Also see [4, 11, 12].
Most researchers in judgmental adjustment agree that, if the contextual information
used is reliable, the performance of judgmental adjustments will be better than that of
judgmental adjustments using unreliable one [5, 13, 14]. Hence, the eliciting of contextual
information from experts using quantitative techniques such as Delphi and cross impact
analysis [15], the screening [4], and classification such as the use of 5 structured types of
causal forces proposed by [16], as well as processing, such as [17, 18], among many others
advocating the use of decomposition method, of relevant contextual information, all are
important aspects of judgmental adjustment.
However, judgmental adjustments still have all kinds of bias like cognitive bias,
double-counting bias, political bias, and so on inherent in judgmental forecasting and still
have the issue of inconsistency [19–21].
The objective of this study focuses on proposing an adjustment mechanism capable
of handling and reflecting, without subjective judgement, detailed changes anticipated in the
forecasting horizon but could not be handled by the regression model alone, which
incorporating all the critical variables estimated with GA in such a way as to be more
realistically conformed with the real world. Thus, this adjustment mechanism, a natural
extension of the model, consisting of seasonal index realignment and proportional
adjustment in a set of equations, is able to make appropriate adjustments consistently and
effectively improving the forecasting accuracy of initial forecasts of the model compared
favorably to other alternative like Box-Jenkins ARIMA [22-23] with adjustment.
The remainder of this paper is organized as follows. Section 2 describes our
forecasting model in detail, including formulation of a regression model, a brief introduction
to the feature and process of GA in model fitting, subsequent model checking, and the
process of e-composition, as well as the adjusting mechanism consisting of seasonal index
adjustment and proportional adjustment, as well as a combination of the former two. Section
3 portrays the background and design of our empirical research. Section 4 depicts the
empirical results of model fitting and model checking, and a comparative analysis of
forecasting results of various adjustment methods assessed with MAPE on per item basis,
percentage of correct direction adjustment, and IMP on per adjustment basis, from the
perspective of adjustment size and lead time, as well as an illustration with two graphical
examples, and discussions. Finally, a conclusion is drawn in section 5.

2. The forecasting model

2.1. Formulation of a regression model

The first objective in our forecasting model involves decomposing the promotional
sales of products of a company into simple components easy to handle. Eq. (1) of our
regression model is motivated by Dick R. Wittink et al.’s analytical models in [24-26]. The
model can be formulated as

( ) ∏μ
n o
θ it
S it = λit Pit / Pi ∏w ε it
, ∀ t ∈Q
D l it H rit
l it rit (1)
l =1 r =1

464
 Quantitative Methods Inquires

Where, i denotes an item number, i = 1,2,3….,I; t denotes specific number of period

referenced, 1 ≤ t ≤ T. T is the total number of normal periods. While I is the total
number of items involved.
Q denotes the set of referenced periods.
Z denotes the set of periods to be forecasted.
Sit is the total unit sales of the item i in period t under a retailer, for weekly sales, t actually
represents a certain week in the referenced periods.
λit denotes the normal unit sales (base sale) of the item i in period t without any promotion
under a retailer.
Pi is the list price of item i.
Pit is the discount price of item i during period t under a retailer.
θ it denotes the coefficient of price elasticity of item i during period t under a retailer.
D denotes an indicator parameter(or dummy variable) of non-price promotion mix.
Dl it is the l-th component of a vector of n indicator parameters of non-price promotion mix

(D1 it , D2 it , …, Dn it ) of item i in period t. D l it = 1 denotes a promotion mix of type l

arises, the default value of D l it = 0.

μl it denotes the non-price promotion effect parameter of type l non-price promotion mix

( Dl it ), a combination of certain non-price promotion activities, of item i during normal

period t under a retailer.
H denotes an indicator parameter of holiday.
Hrit is the r-th component of a vector of o indicator parameters of holiday (H1it , H2it , … , Hoit)
to indicate whether there is any holiday(s) in a certain period t or not. Hrit = 1 denotes a
holiday of type r arises, the default value of Hrit = 0.
w rit denotes holiday effect parameter of holiday type r in period t of item i.
ε it denotes the residual error.

Besides, additional notations listed below may be helpful in the following sections.
ϕ denotes the weekend effect, which is derived via GA based on data in mixed periods.
d(t1) denotes the length of sub-period t1 of t, t ∈ Z . 0 ≤ d(t1) ≤ 7.
d(t2) denotes the length of sub-period t2 of t, t ∈ Z . 0 ≤ d(t2) ≤ 7. d(t1)+d(t2)= d(t), because
in a week there are at most two different kinds of promotion mixes held.
δ denotes the duration of the weekend.
δ t1 denotes the duration of weekend covered by sub-period t1.

Take natural logarithm in both sides of Eq. (1), we get the following:

( )
n o
ln S it = ln λit + θ it ln Pit / Pi + ∑ Dlit ln μ lit + ∑ H rit ln ω rit + ε it , ∀t ∈ Q (2)
l =1 r =1

465
 Quantitative Methods Inquires

Thus, a nonlinear model like Eq. (1) is transformed to a linear regression model
[27-28], which is the underlying model to conduct model fitting and model checking in this
study.

2.2. Model fitting--parameter estimation with GA

To take into account of all the influential and sporadic cue factors in various sub-
periods of the training period, the number of variables may amount to such a quantity that
conventional parameters estimation method like ordinary least square, maximum likelihood
method, and so on may become incompetent, due to the issue of collinearity [29-30],
insignificant parameters, [4, 31] or small sample size. Therefore, in this study we use a
customized genetic algorithm (GA) which could estimate parameters effectively and
efficiently [32].

2.2.1. Features and procedures of GA in this study

GA simulates Darwin’s biological evolution through stochastic crossover and
mutation by selecting encoded individuals (solutions) in the population with higher fitness via
a fitness function to generate population of individuals (reproduction) more fitted to the
environment (better solutions) from generation to generation [33-35].
The initial population is randomly created in the encoded form of a binary matrix,
there are pop rows, each row of binary string in the matrix is an individual (solution) which
encompasses β chromosomes, each chromosome, representing a parameter, is composed of

γ genes, while each gene is represented by a binary code.

Each individual is evaluated by the fitness function, check Eq. (3), in each
generation, the best α % ( 1 ≦ α ≦ 6) of the population are kept as elites to the next
generation, the remaining of the population are created by a randomly selected pairs of
individuals conducting a multi-point crossover [36], n + o + 2 points in total, for each one of
each pair to reproduce offspring, forming a random recombination of individuals’
ingredients of genes, to search new solution space and possibly better solution.
After that, a one-bit mutation is performed [37], with a view to creating new pieces
of gene originally not possessed by members of the population, through randomly selected
genes within each individual, this occasional random change in genes could open the door
to new possibilities of better solutions. Afterwards, each encoded individual in the population
is decoded back to a string of real numbers of parameters, and each individual is evaluated
by the fitness function…, the iterative process goes on and on until a termination condition is
met.
In this study, parameters like crossover probability (Pc) and mutation probability (Pm)
of GA are designed to vary with the number of generations processed or others, such as the
minimum level of moving average percent of improvement (MAPI) in fitness function value
within certain number of generations, to keep proper diversity of the population, while
retaining the convergence capability, to circumvent getting stuck too early at local solutions
in its search process and derive satisfying results [38-39].
In estimating parameters of complicated multivariate nonlinear models, GA is
generally considered to be better than other alternatives such as nonlinear least square,

466
 Quantitative Methods Inquires

maximum likelihood estimation, and so on, due to its parallel search capability [40-41], even
based on small size dataset, it is capable of deriving satisfying results.

The fitness function of GA may be formulated as

T
~
FVi = MAPEi = (∑ ln S it − ln S it / ln S it ) / T , ∀t ∈ Q (3)
t =1

~
Where the term ln S it − ln S it is the absolute value of difference between natural

logarithm of the actual sales volume (ln S it ) of the i-th item and natural logarithm of the
~
estimated sales volume (ln S it ) of the same item in period t. T denotes the number of normal
periods. The objective of GA is to find a solution with the minimal MAPEi. The smallest MAPEi
found is updated once a smaller one is found in the solution search process. After model
fitting, every effect parameter in Eq. (2) is derived in real value.

2.3. Model checking

In this section, a regression diagnostics focused on normality and independence is
performed to see if critical assumptions of linear regression are violated, based on Eq. (2). If
these assumptions are severely violated, particularly if collinearity arises among predictor
variables, bias may be a serious issue in model fitting or even in model specification.
Normality test is conducted through One-Sample Kolmogorov-Smirnov test [42],
and Q-Q plot [43]. Independence test in this study consists of two parts, namely, multi-
collinearity test and autocorrelation test. The former is performed via condition index,
whereas the latter is performed via ACF checking [44].

2.4 The re-composition of effect parameters

As the cycle length of CPG industry is about 52 weeks long, let t’ = t + 52,
denoting the corresponding week to be forecasted in a new year. A modified naïve sales
forecasting method considering cycle length to forecast unit sales of item i of period t’ in a
new year, see Williams (1987), according to sales data of week t in the referenced year,
would be

( ) ∏μ
n o
θ it '
Sˆit ' = η iπ it ln Pit ' / Pi ∏ω t ' = t + 52, ∀t '∈ Z .
Dlit ' H rit
it ' rit , (4)
l =1 r =1

Where,η i denotes the average normal sale of item i across referenced

periods. π it denotes the seasonal index of item i in period t, and Z denotes the set of periods
to be forecasted.
So far, all the parameters in Eq. (4) are already derived via GA. Let e1it’ denotes the
price effect multiplier of item i in forecasting period t’, e2 it’ and e3 it’ denote the effect
multiplier of a non-price promotion mix and a specific holiday effect, respectively.. In each
group of indicator parameters at most one condition will arise in each period. We get

467
 Quantitative Methods Inquires

Sˆit ' = ηiπ it e1it e2it e3it , t ' = t + 52, ∀t ' ∈ Z . (5)

In its re-composed form, Eq. (5) can be used to forecast weekly unit sales. Actually,
parameters estimated through GA, based on observations in the training periods, can be
recombined as Eq. (5) in responding to expected promotional campaigns in the forecasting
horizon specified in the promotion proposals to perform out of sample forecasting without
any adjustment in the following empirical study.

2.5. The adjusting mechanism of this study

The mechanism of this study stresses that adjustments of forecasting are based on
the anticipated changes of the context of promotions and holidays in the forecasting periods,
which can not be handled by the regression model alone, in this study a set of equations are
formulated to do this job, they are natural extension of the model. The objective is to
improve the performance of the model, making our final forecasts more closely reflect these
changes in prospect.

2.5.1. Seasonal index adjustment (SIA)

Based on domain knowledge, sales volume of the last week or average sales
volume of the last few weeks (adjusted with calendar effect) in the reference periods is a
better predictor to sales of the first few weeks in the forecasting periods next year than sales
of the same weeks in the referenced periods in Taiwan, check Figures. 1-2. Thus, the
corresponding formula can be modified from Eq. (5) and formulated as

Sˆit ' = η iπ it ( t =Ω ) e1it ' e2 it ' e3it ' e4 it '( t '=t +52− ws ) , ∀t '∈ Z (6)

Eq. (6) is an example of modified Naïve method used for multiple-step out-of-
sample forecasting considering cycle length which is a year. In which, η iπ it (t =Ω ) stands for
normal (base) sales of the last week(s) in the training periods and is the most recent related
data available. While e4 it ' stands for pre LNY (Lunar New Year) effect arises annually in a
period of around 4 weeks right before LNY. In this period, sales volumes are usually much
higher than usual even without any promotion. Since pre LNY effect has not been
incorporated as a variable into our regression model for parsimonious purpose, it will
dominate the seasonal indices in corresponding periods, hence, it’s quite intuitive to use
these indices as proxy variable of pre LNY effect, denoted as e4 it ' . Because there is usually a
time shift of the timing of LNY from year to year, to forecast unit sales of weeks after LNY in
next year, the week number referenced corresponding to the week in the forecasting horizon
has to be adjusted.
Let LNY(t’) denote the week number of LNY in the year to be forecasted, LNY(t)
denote the week number of LNY in the referenced year, then, let ws = LNY(t’) – LNY(t)
represents the number of weeks shifted between two different years as the week of LNY is
concerned. If ws > 0, it means the sequence of week of LNY(t’) in the forecasting year will be
ws weeks later than that of LNY(t) in the referenced year; on the other hand, if ws < 0, it
means the sequence of week of LNY(t’) in the forecasting year will be ws weeks earlier than

468
 Quantitative Methods Inquires

that of LNY(t) in the referenced year. Therefore, the most right term in Eq. (6) e4 it ' will be

replaced by π i ( t − ws ) . Thus Eq. (6) will become

Sˆ it ' = η i π it ( t = Ω ) e1 it ' e 2 it ' e3it 'π i ( t − ws ) , ∀t '∈ Z (7)

sales of week 1 in 2007

sales of week 1 in 2007

1200 1200
y = 0.9417x - 27.447 1000 y = 0.7388x + 103.83
1000
R2 = 0.9556 R2 = 0.6494
800 800
600
600
400
400
200
200
0
0
0 500 1000 1500
0 500 1000 1500
sales of week 1 in 2006
sales (adjusted) of week 52 in 2006

Figure 1. Regression of week 52’s sales Figure 2. Regression of week 1’s sales
(2006) to sales of week 1 in 2007 (2006) to sales of week 1 in 2007

We also find that sales of the same weekly order as that in a different year after
LNY are more aligned than weekly sales of the same ordinary sequential order between
different calendar years, check Figures 3-4, in which R2 from sales of weeks after LNY in
2007 regressed against those of weeks after LNY of 2006 is 0.8, much better than the
ordinary week n corresponding to the same week n, n = 1,2,..,5, regression between
different years, which only has a R2 of 0.628, please check Figure 4.
sales of w9-w13 2007

sales of w6-w10 2007

1,000 2500

800 y = 0.7504x + 29.13 2000

R 2 = 0.8007 y = 1.2861x - 14.397
600 1500
R2 = 0.6283
400 1000

200 500

0 0

0 200 400 600 800 1,000 1,200 0 500 1000 1500

sales of w6-w10 2006 sales of w6-w10 2006

Figure 3. Regression of sales of w6-w10 Figure 4. Regression of sales of w6-w10

2006 to that of w9-w13 in 2007 2006 to that of w6-w10 in 2007

Based on the finding mentioned above, to forecast the sales of weeks after the
week of LNY in a new year, denoted as Sˆit ' , t ' > LNY (t ' ) , Eq. (8) can be of use, which is
modified from Eq. (7) :

Sˆit ' = η iπ i ( t − ws ) e1it ' e2 it ' e3 it ' , t ' > LNY (t ' ) , ∀t '∈ Z (8)

469
 Quantitative Methods Inquires

2.5.2. Proportional adjustment (PA)

Quite often, in the week of LNY, there is a small part of pre LNY present prior to
the eve of LNY, or the last week of pre LNY is mixed with a small part of LNY in the
referenced period, but the condition of the corresponding week in forecasting horizon is
different, in these cases, to get a proper estimation of these effect multipliers of calendar
effect in the forecasting periods, we must get them restored to regular ones ( a whole week
only covered by purely one kind of holiday related effect like pre LNY effect or holiday effect
of LNY) first, and then proceed to calculate the changed mixed effect in the forecasting
period.
The adjusting equations are used to calculate the mixed effect of pre LNY and LNY
present in the same week:

(d (t1 ) − d (δ t 1 ) + d (δ t 1 )ϕ )e4it1 + (d (t 2 ) − d (δ t ) + d (δ t 1 ) + (d (δ t ) − d (δ t 1 ))ϕ )e*3it2

e3 i t = , ∀t ∈ (Q ∪ R) (9)
d (t ) + d (δ t )(ϕ − 1)

In Eq. (9), the part ( d (t1 ) − d (δ t ) + d (δ t )ϕ )e4it1 represents the sum of the effect
1 1

of the last week in pre LNY in the duration of weekdays covered by sub-period t1 and the
effect of the last week in pre LNY in the duration of weekend covered by sub-period t1 times
the weekend effect. While the term ( d (t 2 ) − d (δ t ) + d (δ t ) + ( d (δ t ) − d (δ t ))ϕ )e
*
3it 2 stands
1 1

for the sum of the effect of regular LNY in the duration of weekdays covered with sub-period
t2 in the referenced period and the effect of regular LNY in the duration of weekend covered
by sub-period t2 in the referenced period times the weekend effect. Every parameter in Eq.
(9) except e*3it is known, so e*3it can be obtained. Note that the daily effect of each normal
weekday in a week is assumed to be 1. Eq. (9) actually is a daily effect weighted average
formula of mixed weekly effect of pre LNY and LNY present in the same week.
It follows that the mixed effect in the forecasting period ( e3 i t ) could be computed
1'

through the following formula:

(d (t1 ' ) − d (δ t 1 ' ) + d (δ t 1 ' )ϕ )e* 3it1 ' + (d (t 2 ' ) − d (δ t ) + d (δ t 1 ' ) + (d (δ t ) − d (δ t 1 ' ))ϕ )erit2 '
e3 i t ' = , ∀t '∈ Z (10)
d (t ) + d (δ t )(ϕ − 1)

Where, erit2’ may be effect of the last week in pre LNY or just base effect equal to 1.
In Eq. (10), the part ( d (t1 ' ) − d (δ t ' ) + d (δ t ' )ϕ )e
*
3it1 ' stands for the sum of the effect of the
1 1

regular LNY in the duration of sub-period t1’ in the weekdays in the forecasting period and
the effect of the regular LNY in the duration of sub-period t1’ in the weekend times the
weekend effect. While the part of (d (t 2 ' ) − d (δ t ) + d (δ t 1 ' ) + (d (δ t ) − d (δ t 1 ' ))ϕ )erit2 '
represents the sum of the effect of the last week in pre LNY in the duration of sub-period t2’in
the weekdays in the forecasting period and the effect of the last week in pre LNY in the
duration of sub-period t2’ in the weekend times the weekend effect.
In the same token, regular effect of the last week in pre LNY (e*4it1’) in the
referenced periods can be derived via Eq.(11). Then, the mixed effect of e4it1’ could be
obtained with Eq. (12):

470
 Quantitative Methods Inquires

(d (t1 ) − d (δ t 1 ) + d (δ t 1 )ϕ )e*4 it1 + (d (t2 ) − d (δ t ) + d (δ t 1 ) + ( d (δ t ) − d (δ t 1 ))ϕ )e3it2

e4it = , ∀t ∈ ( Q ∪ R ) (11)
d (t ) + d (δ t )(ϕ − 1)

(d (t1 ' ) − d (δ t 1 ' ) + d (δ t 1 ' )ϕ )e* 4it1 ' + (d (t 2 ' ) − d (δ t ' ) + d (δ t 1 ' ) + (d (δ t ' ) − d (δ t 1 ' ))ϕ )e3it2 '
e4 i t ' = , ∀t '∈ Z (12)
d (t ) + d (δ t ' )(ϕ − 1)

2.5.3. Total adjustment (TA)

As the combination of both SIA and PA, TA is the most comprehensive adjustment
in this study.

3. Empirical Study

3.1. The background of empirical study

This study has a focus on the adjustment of model-based forecast of weekly unit
sales of several series of CPG products, manufactured by Company A, under retailer B.
Company A is a leading manufacturer specialized in dehumidifier and deodorizer products
in Taiwan. While retailer B is an international outlet of DIY products.
A sales data set of 10 items in 2007, aggregated from retailer B’s outlets, coupled
with price promotion, non-price promotion data, as well as promotion proposals, which were
set up in 2007, of the first 4 months in 2008, are used to conduct our empirical study. The
training dataset covers two periods, the first period covers the whole year of 2007, the
dataset of this period can be denoted as sample A, forecasting horizon is the first 6 weeks of
2008. The second period ranges from the beginning of 2007 to the 10th week of 2008, the
training data of this period can be denoted as sample B, forecasting horizon ranges from
11th week to 16th week of 2008.
The underlying equation used in model fitting was Eq. (2), all the effect parameters
in Eq. (2) were estimated through GA with objective function set as Eq. (3) and constraints set
realistically from contextual knowledge, such as price elasticity parameter to be in the range
of [0, -8], non-price promotion effect multiplier to be in the range of [1, 5], holiday effect
multiplier to be in the range of [1, 2]. Besides, the number of types of non-price promotion
mixes n in Eq. (2) was set to be 7, therefore, there are about 7 types of different combination
of promotion activities across seasons, and the number of holiday type o was set to be 4,
which means there are about 4 different types (according to duration of holiday) of holidays
each year, to reflect the actual business settings. GA programs were run with Matlab 7.1.
Effect parameters estimated with GA and mixed effect parameters reassessed in the mixed
periods on both sample A and sample B were recomposed according to the expected
variations of promotions in the promotion proposal as initial forecasts without any
adjustment. The multiple-step out-of-sample forecasts with ARIMA were run with SPSS 13. In
busy season, because of intensive promotion campaigns, ARIMA tends to underestimate unit
sales, the rule of adjustment for forecasts of ARIMA can be formulated as

ARIMA ad = forecast of ARIMA * 1.2 (13)

However, in time of off season, ARIMA tends to overestimate unit sales, the rule of
adjustment can be formulated as

471
 Quantitative Methods Inquires

ARIMA ad = forecast of ARIMA * 0.8 (14)

Thus, the performance of various adjustment methods in this study can be

compared with their counterpart of ARIMA.

3.2. The design of empirical study

In order to take both the busy season and off season into account to have a proper
assessment of the performance of different adjustment methods, the forecasting horizon is
designed to consist of two periods of equal duration, the first period includes the first 6
weeks of 2008 which covers the busiest season, ie, the LNY season in Taiwan, and can be
denoted as busy season, while the second one starts from the 11th week and ends at the16th
week of 2008, which is one of the off seasons in the same year and can be denoted as off
season.
As the forecasting target is concerned, 10 items of products were selected to
conduct our empirical research. The relevant prices and promotion activities can be found in
promotion proposals which actually are the source of anticipated variations in promotions.
Another source of anticipated variations in calendar effects is the calendar.
To properly evaluate the performance of various methods in adjusting original
forecasts of the model, which can be denoted as NA, made by regression model, SIA was
performed first to adjust NA, followed by PA to adjust the same NA. Then, the combination
of SIA and PA were used to adjust NA, denoted as TA. The busy season was the first
forecasting horizon, and off season was the second forecasting horizon. In addition, for the
purpose of comparative reference, forecasts with Box and Jenkins ARIMA were derived,
adjustments of forecasts from ARIMA were derived with Eq. 15 in busy season, and Eq. 16 in
off season, respectively.

4. Empirical results

4.1. The results of model fitting

The estimated error in terms of MAPE in general is below 3%, except item 10. Most
parameters derived are consistent with our expectations, such as the effect parameters of μ1

to μ3 are increasingly bigger from 2.391 to 2.992 for sample A and from 2.382 to 2.848 for
sample B, respectively, because more effort and expenditure are made for the bigger
number type of promotion therein, and μ5 is larger than μ4 because non-price promotion
type 5 employs direct mail in addition to what type 4 has.

4.2. The results of model checking

The normality test, consisting of one-sample Kolmogorov-Smirnov test and Q-Q
plot, in which, this model passed the test with data from sample A or sample B without
problem based on standardized error term ε it in Eq. (2) and natural logarithm of predicted
~
weekly unit sales denoted as ln S it in Eq. (3). However, in independence test, the measures
of condition index and results of ACF showed complex but interesting consequences in both
samples, in which 5 out of 10 items have autocorrelation problems for sample B. As for
sample A, there are 2 items have the same kind of problems, however. As collinearity is
concerned, according to [44], if the condition index is above 10 and below 30, there may

472
 Quantitative Methods Inquires

have a minor problem of collinearity, if condition index is above 30 and below 100, there
may have moderate to severe collinearity issue. In our empirical study, 4 out of 10 items
may have moderate to severe collinearity problems, 5 items may have severe collinearity
problem for sample B, the condition looks similar for sample A, check Table A1 in Appendix
A. If model parameters are estimated by OLS, it is quite possible to have serious bias issues.

4.3. Comparing and analyzing results from various kinds

of forecasting adjustment methods
The performance of weekly sales forecasting adjustment from various methods in
terms of MAPE can be displayed in Table 1 and Table 2. Each cell with negative adjustment
performance is in bold face. Among these adjustment methods, without taking advantage of
any adjustment, the MAPE of sales forecasting with the regression model, that is NA, in
average, is 17.96% and 37.06% in busy season and off season, respectively.
If forecasts are adjusted with SIA (seasonal index adjustment), the average
performance across items in terms of MAPE is 19.51% for busy season, which seems a little
worse than NA, check Table 1. However, for off season, the average figure of SIA is 24.77%,
a significant improvement of 33.16% of initial MAPE, check Table 2. There are 5 items
improved out of 10 because of SIA for busy season, while there are 7 items get improved
due to SIA’s contribution for off season. The relatively poor performance of SIA in busy
season may be attributed to the already good performance of NA compared to that of
ARIMA without adjustment.
If adjustment is conducted with PA (proportional adjustment of mixed effect in
mixed periods) in busy season, we see an improvement from 17.96% to 14.98%, a 16.59%
improvement in average. The number of items with negative results is reduced to 3 also,
check Table 1. However, for off season, it’s a different story for PA, with MAPE just slightly
reduced from 37.06% to 34.37%, a small improvement of 7.26% overall in average.
Nevertheless, 5 out of 7 adjustments performed gets improvement in MAPE, besides, item 3,
8, and 10 didn’t perform any PA adjustment, therefore, their MAPE are the same as that of
NA, and almost only 1 out of 6 weeks needs to get adjustment with PA in off season, their
contribution to the improvement of MAPE is therefore trivial, check table 2.

Table 1. Comparison of the accuracy of various forecast adjustment

methods in busy season 2008
MAPE of various forecast adjustment methods
item NA SIA PA TA ARIMA* ARIMA ad
1 22.65% 18.18% 16.30% 11.87% 39.00% 60.14%
2 9.00% 16.53% 7.55% 15.12% 10.84% 24.41%
3 28.91% 28.67% 17.48% 20.91% 52.29% 42.75%
4 24.93% 21.90% 15.76% 18.66% 41.09% 29.31%
5 9.88% 14.15% 11.40% 10.20% 16.02% 20.01%
6 21.30% 19.43% 21.82% 18.75% 37.46% 30.14%
7 13.80% 23.77% 13.17% 25.62% 21.19% 37.57%
8 8.95% 18.11% 7.81% 15.65% 27.79% 20.32%
9 23.14% 17.34% 24.10% 15.97% 39.04% 26.85%
10 16.99% 17.01% 14.45% 16.05% 23.23% 19.35%
AVG 17.96% 19.51% 14.98% 16.88% 30.80% 31.09%
Note: forecasting with ARIMA (1, 1, 1)

473
 Quantitative Methods Inquires

If TA (total adjustment) is performed, since it combines both SIA and PA, for busy
season, average MAPE reduced from 17.96% to 16.88%, the reduction of MAPE amounts to
an average of 6.01% in improvement. For off season, the improvement is even more
significant, the MAPE of NA improved from an average of 37.06% to an average of 22.78%,
about 38.51% improvement over MAPE of initial forecasts. Note that if both SIA and PA have
positive contribution in improving forecast accuracy, TA can be very effective, check item 1
and item 4 in Table 1, also item 1, 4, 5, 6, and 7 in Table 2.
If the adjustment of ARIMA is of concern, 6 out of 10 items get improved in busy
season, check Table 1, but average percent of improvement from adjustment is a negative -
0.942%, overall performance of ARIMA adjustment seems to be worse than that of PA and
TA, however, it still is better than SIA in busy season.

Table 2. Comparison of the accuracy of various forecast adjustment methods

in off season 2008

MAPE of various forecast adjustment methods

item NA SIA PA TA ARIMA ARIMA ad
1 33.03% 20.93% 23.27% 12.12% 46.95% 18.95%
2 12.77% 33.43% 16.58% 36.78% 21.23% 20.80%
3 36.16% 12.37% 36.16% 12.37% 40.07% 52.05%
4 39.28% 23.41% 38.52% 24.32% 38.86% 17.87%
5 90.76% 46.36% 85.31% 41.77% 29.58% 23.08%
6 59.41% 8.83% 56.82% 10.43% 15.94% 26.55%
7 37.50% 32.12% 25.04% 19.38% 32.02% 28.37%
8 24.50% 19.55% 24.50% 19.55% 16.56% 8.38%
9 17.09% 19.03% 17.36% 19.44% 11.51% 11.93%
10 20.13% 31.63% 20.13% 31.63% 34.87% 47.90%
AVG 37.06% 24.77% 34.37% 22.78% 28.76% 25.59%

In off season, the adjustment of ARIMA displays a performance obviously better

than its counterpart in busy season, the number of items get improved in MAPE remains the
same as that in busy season. However, overall percentage of improvement from adjustment
amounts to 11.02%, a performance much better than its counterpart in busy season.

4.3.1. Analysis of various adjustment methods from the

perspective of adjustment size
In this subsection, the percentage of correct direction in adjustments over total
adjustments is used to measure the performance of various adjustment methods. Whether
the direction is correct or not depends on the comparison between initial forecast and the
actual sale, if initial forecast is under-forecast, the correct direction of adjustment should be
adjusted upwards, regardless of adjustment size. On the other hand, if initial forecast is
over-forecast, the correct direction of adjustment should be adjusted downwards, regardless
of adjustment size. However, if the initial forecast is within the range of [actual unit sales -
3%*actual unit sales, actual unit sales + 3%*actual unit sales], any subsequent adjustment
with result less than or equal to initial over-forecast or any adjustment with result more than
or equal to the under-forecast is perceived as adjustment in the correct direction.
In this study, any adjustment with result within less than 10% range of the initial
forecast, regardless of adjustment direction, is regarded as small adjustment, otherwise, it’s
a large adjustment. In Table 3, all the small adjustment, in both busy season and off season,

474
 Quantitative Methods Inquires

has the ratio of adjustment with correct direction regardless of adjustment method, to be less
than 60%.
On the other hand, large size adjustments seem to have a much more consistent
and better performance in average than that of small ones, with the average correct ratio at
least over 70% in off season except ARIMA adjustment, particularly, in busy season, an
average ratio of correct direction around 86% is recorded for methods proposed in this
study, whereas for adjustment of ARIMA, the percentage is 73.33%, still is not bad. This
result is not surprising, in the literature, there are considerable similar evidences (Fildes and
Goodwin, 2007; Syntetos et al., 2009).
Among three adjustment methods proposed in this study, SIA seems to have the
best performance in terms of the ratio of correct direction adjustment in both small and large
adjustments in off season, the overall ratio of correct direction adjustment on the basis of per
adjustment is about 80%, but it does not necessarily mean that SIA provides the most
positive contribution to improvement of forecast accuracy, because in Table 4, whether the
adjustment is over-adjusted or not is not taken into account. Table 4 may offer some
remedies in this regard. However, in busy season, PA seems to be the winner, with an
overall ratio of correct direction adjustment close to 80%, whether it offers the most positive
contribution to forecast accuracy improvement or not, still have to be crosschecked with
other criteria like IMP in Table 4 to have an adequate assessment.
A measure called IMP, which can be used to evaluate the adjustment improvement, may
be formulated as

IMP = APEini - APEad (15)

Where, APE denotes absolute percentage error, APEini denotes APE of initial
forecast, while APEad denotes APE after adjustment.
In Table 4, with the only exception of SIA applied in busy season, large adjustments
consistently and significantly outperform small adjustments in terms of IMP, regardless of the
adjustment method. The only exception is SIA which implies that most large size SIA
adjustments with correct direction in Table 3 are actually over-adjusted. Note that in busy
season, all three adjustment methods using small adjustment, the average IMP are all
negative, among them, more than half of small adjustments made by SIA and TA are in
correct direction, this means that there are serious issue of over-adjustment in the small size
adjustments of these two adjustment methods.

Table 3. Comparing the performance concerning direction of adjustment of various adjust

busy season 2008 off season 2008
% of correct % of correct % of correct % of correct
AD ratio of ratio of ratio of ratio of
direction in direction in direction in direction in
method small ad large ad small ad large ad
small ad large ad small ad large ad
SIA 30/60 53.33% 30/60 83.33% 14/55 57.14% 41/55 87.80%
PA 5/20 40.00% 15/20 93.33% 1/8 0.00% 7/8 71.43%
TA 25/60 56.00% 35/60 82.86% 12/60 50.00% 48/60 79.17%
ARIMA ad 0/60 -- 60/60 73.33% 0/60 -- 60/60 58.33%

475
 Quantitative Methods Inquires

Table 4. Comparing IMP of various forecast adjustment methods

in either small or large djustments
busy season 2008 off season 2008
AD avg IMP from avg IMP from avg IMP from avg IMP from
method small adjustments large adjustments small adjustments large adjustments
SIA -0.21% -5.63% 0.66% 10.8%
PA -3.86% 14.37% -1.7% 37.3%
TA -0.395% 4.44% 1.37% 28.8%
ARIMA ad -- -1.036% -- 2.89%

As large size adjustments are concerned, in Table 3, even though PA is not the best
performer in terms of percent of correct direction adjustment, however, in Table 4, it does
have the best performance in terms of IMP, this implies that the number of over-adjustments
of PA is the smallest. In busy season, PA still is the best performer in terms of IMP, it is the
best one even from the viewpoint of percent of correct direction adjustment, crosscheck with
Table 1, obviously, it provides the most consistent and the largest contribution to the
improvement of forecast accuracy in busy season among various methods. However, in off
season, because of the relatively less frequency of PA adjustments made, even though it still
offers the best performance in terms of IMP, check Table 4, its overall contribution to
improvement of forecast accuracy in terms of MAPE is not impressive.
On the other hand, the performance of TA in terms of IMP in off season though is
not the best among these methods, due to its highest ratio of large size adjustment, check
Table 3, TA still provides the most positive contribution to improvement of forecast accuracy
in terms of MAPE, check Table 2.

4.3.2. Analysis of various adjustment methods from the perspective of lead time
In Figure 6, each forecasting horizon is divided into two parts, namely, the first 3
weeks and the second 3 weeks in both busy season and off season. Obviously, in busy
season, the performance of various adjustment methods is relatively more stable than that of
its counterpart in off season. Among different adjustment methods, PA seems to have the
best performance in terms of IMP across different seasons, TA ranked second, and SIA still is
the worst performer. Note that PA is only conducted in the second half of the forecasting
horizon, check Figure 5, unlike other methods, it appears as a point in each season.
The performance of adjustment of ARIMA looks not so stable in different lead time
of busy season, however, in off season, its performance is parallel to others in shape but
obviously inferior to others.

4.4. Illustration of typical adjustment of different

methods with two examples
To further present the detailed results of each adjustment method compared to
initial forecast and Box and Jenkins ARIMA’s forecasts and their simple adjustments as a
yardstick, two graphs are drawn, see Figures 6-7. Note that to expose the detailed results of
adjustments more clearly, all data from week 11 to week 50 of 2007 are cut off. The
forecasting horizon is the first 6 weeks of 2008 (the busy season) in Figure 6.
Note that the original forecasts of the regression model are expressed in purple
square dots in purple line, after adjustment of SIA, which are expressed in yellow triangle
dots with yellow line, in weeks 3-4, the direction of adjustment are not correct, therefore, SIA

476
 Quantitative Methods Inquires

actually made obvious negative improvements. There is no adjustment made in the first 4
weeks for PA, However, in weeks 5-6, due to excellent adjustments of PA, which are
expressed in empty blue diamonds with bold blue line, the forecast points are moved much
closer to the actual sales, check TA in weeks 5-6 in Figure 6. Also note that down in the
bottom are points formed by ARIMA, after 20% upward adjustment, these points are moved
closer to the reality.
To illustrate the adjustment performance of different adjustment methods and
initial forecasts as well as forecasts and adjustment of ARIMA in a bigger scope, the same
item is used in Figure 7, in which, initial forecasts of the model are expressed in purple dots
which are not close to the reality, after adjustments of SIA, check yellow triangle dots in
weeks 11-16. In weeks 15-16, owing to the adjustment of PA, the forecasts are moved much
closer to the reality. Because of the relative poor performance of NA in the first 4 weeks in
Figure 7, and PA doesn’t make any adjustment in this period, its overall performance is
unlike its performance in Figure 6. TA, due to its combination of SIA and PA, particularly SIA,
which has a good performance and therefore push overall TA closer to the reality.
The points of ARIMA stays in relative high positions from week 11 to week 16 and
are also forming the very one most far away from the reality, after a 20% downward
adjustment, they are moved much closer to the actual sales in average, check the bold green
line with empty circles in Figure 7.

30
27.2
25 22.3
21.53
average IMP (%)

20
SIA
15
11.5 PA
10 10.1 10.3
5.25 TA
5.54 6.53
5 ARIMA ad
-0.83 3.69
0 -1.07
-4.18 -4.51
-5
-10
first 3 weeks second 3 weeks first 3 weeks second 3 weeks

busy season off season

Figure 5. Comparison of average IMP of various adjustment methods on different lead times

4.5. Discussions
From the above analysis and explanations, on a per adjustment basis, PA offers the
most improvement in terms of both MAPE and IMP in busy season, check Table 1 & 4, it also
has the highest percentage in correct direction adjustment in busy season. In off season,
since it is rarely used (the mixed effect condition is rare in comparison), its total contribution
is not impressive.
TA, on the other hand, is more comprehensive in off season, and provides the most
contribution in improving MAPE, even though in terms of percentage of correct direction
adjustment, it is not the best performer. Crosscheck Table 3 and Table 4, it is easy to see
that, the correct direction is the prerequisite for an adjustment of any kind to improve
forecasting accuracy, but due to the issue of over-adjustment, many correct direction
adjustments still have negative contributions to forecasting accuracy.

477
 Quantitative Methods Inquires

350
forecast of unit sales

300

250

200

150

100
1 2 3 4 5 6 7 8 9 10 51 52 1 2 3 4 5 6
WEEK NO.

Actual NA SIA PA TA ARIMA ARIMA ad

Figure 6. Comparison of adjustment performance in busy season 2008 with

different adjustment methods on item 4.

240
220
forecast of unit sales

200
180
160
140
120
100
80
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
WEEK NO.

Actual NA SIA PA TA ARIMA ARIMA ad

Figure 7. Comparison of adjustment performance in off season 2008 with

different adjustment methods on item 4.

As adjustment of ARIMA is concerned, its performance in terms of MAPE is negative

in average in busy season, nevertheless, in off season, the average MAPE after adjustment is
a not bad 25.59%, a performance quite close to that of TA in the same season. Therefore, if
efficiency is an important issue in forecasting, ARIMA with adjustment is a good alternative in
off season. Otherwise, TA is the best tool of adjustment in off season. In busy season, PA is
the best choice in forecast adjustment, due to its significantly much better performance.
Besides, with the only exception of SIA in busy season, if the performance is
measured in terms of both percentage of correct direction adjustment and IMP in average,
large size adjustment has a significant advantage over small size adjustment, regardless of
season. Our model assumes that there will be no big difference between actual promotion
activities and those specified in promotion proposals in forecast horizon, if this is not true,

478
 Quantitative Methods Inquires

there will be larger MAPE incurred in the original forecasts and various types of forecast
adjustments. The relatively more accurate performance of original forecasts in busy season
than in off season may due to the fact that promotion and calendar effects are so strong that
they dominate unit sales in busy season, while in off season, these effects are much less
obvious and much less frequent as in busy season, other factors like seasonal index
realignment, competitors’ actions and so on may have critical impacts on unit sales therein.

5. Conclusions

The forecasting adjustment mechanism proposed in this study only concerning with
realignment of seasonal indices and the anticipated mixed effect of certain variables, such as
the multiplier of the effect of promotion mix, and the multiplier of holiday effect, already
incorporated in the regression model and assessed with GA which is more flexible and is
capable of deriving more realistic coefficient of variables than most other conventional
alternatives. Therefore, adjustment mechanism proposed in this study is a necessary and
natural extension of the regression model. And in the process of forecast adjustment,
subjective judgement based on contextual information is minimized.
Among three adjustment methods embedded in the adjustment mechanism of this
study, SIA focuses on the realignment of seasonal index in forecasting horizon in a different
year than the year of referenced periods, PA provides the necessary reassessment of mixed
effect in mixed periods and is capable of offering the most contribution to the improvement
of forecasting accuracy on per adjustment basis and is also the best performer in average in
busy season. However, in off season, since both SIA and PA provide positive contribution in
improving forecasting accuracy, and TA combines the above two adjustment methods, it
offers the most comprehensive and most reliable adjustment in off season of this study, even
though it’s not necessarily the best performer in improving initial model-based forecast on a
per adjustment basis in average.
In off season, ARIMA with adjustment, which just moves down the original forecast
by 20%, also provides very good forecast accuracy close to that of TA in average, besides,
ARIMA is embedded in most statistical software package and is very handy, in case efficiency
is an important requirement in forecasting, ARIMA with adjustment intuitively is a good
alternative.

References

1. Armstrong, S. Forecasting by extrapolation: Conclusion from 25 years of research,

Interfaces 14, 6, 1984, pp. 52-66
2. Armstrong, S. Long-range forecasting: from crystal ball to computer, New York; Wiley, 1985
3. Armstrong, S., Collopy, F. and Thomas, Y. J. Decomposition by causal forces: a procedure for
forecasting complex time series, International Journal of Forecasting, 21, 2005, pp.
25-36
4. Belsley, D. A., Kuh, E., and Welsch, R. E. Regression Diagnostics. Identifying Influential Data
and Sources of Collinearity, John Wiley & Sons, New York, 1980
5. Belsley, D.A. Assessing the presence of harmful collinearity and other forms of weak data
through a test for signal-to-noise, Journal of Econometrics, 20, 1982, pp. 211-253

479
 Quantitative Methods Inquires

6. Berilant, J., de wet, T., and Goegebeur, Y. A goodness-of-fit statistic for Pareto-type
behavior. Journal of Computational and Applied Mathematics, 186, 1, 2005, pp. 99-
116
7. Blattberg. R. C. and Hoch, S. J. Database models and managerial intuition! 50% model +
50% manager, Management Science, 36. 8, 1990, pp. 887-899
8. Box, G. E. P., Jenkins, G. M., and Reinsel, G. C. Time series analysis: Forecasting and control
(3rd ed.), Englewood Cliffs, NJ7 Prentice Hall, 1994
9. Bunn, D. W. and Salo, A. A. Adjustment of forecasts with model consistent expectations,
International Journal of Forecasting, 12, 1996, pp. 163-170
10. Carroll. R. J. and Ruppert. D. Transformation and Weighting in Regression, Chapman and
Hall, New York, NY, USA, 1998, pp. 115-160
11. Dalrymple, D. J. Sales forecasting practices, results from a United States survey,
International Journal of Forecasting, 3, 1987, pp. 379-391
12. De Gooijer, J. G. and Hyndman, R. J. 25 years of time series forecasting, International Journal
of Forecasting, 22, 2006, pp. 443- 473
13. De Jong, K. A. and Spears, W. M. A formal analysis of the role of multi-point crossover in
genetic algorithms, Annals of Mathematics and Artificial Intelligence 5(1), 1992, pp.
1–26
14. Edmundson, R. H., Decomposition: A strategy for judgmental forecasting, Journal of
Forecasting, 19, pp. 305-314, IEEE Transactions on Evolutionary Computation, 3(2),
1990, pp. 124-141
15. Eiben, A. E., Hinterding, R. and Michalewicz, Z. Parameter control in evolutionary algorithms,
IEEE Transactions on Evolutionary Computation, 3(2), 1999, pp. 124-141
16. Fildes, R. and Goodwin,P. Good and bad judgement in forecasting: Lessons from four
companies, Foresight, Fall, 2007, pp. 5-10
17. Foekens, E. W., Leeßang, P. S. H. and Wittink, D. R.. Varying parameter models to
accommodate dynamic promotion effects, Journal of Econometrics, 89, 1999, pp.
249-268
18. Franses, P. H. and McAleer, M. Testing for unit roots and non-linear transformations, Journal
of Times Series Analysis, 19, 1998, pp. 147-164
19. Goldberg, D. E. Genetic Algorithms, in Search, Optimization & Machine Learning, Addison-
Wesley, Boston, MA, USA, 1989
20. Goldberg, D. E. Simple genetic algorithms and the minimal deceptive problem, in: Davis, L.
(ed.) “Genetic Algorithms and Simulated Annealing”, Hyperion Books, New York, NY,
USA, Morgan Kaufmann, San Francisco, CA, USA, 1987, pp. 74-88
21. Goodwin, P. and Fildes, R. Judgmental forecasts of time series affected by special events:
Does providing a statistical forecast improve accuracy? Journal of Behavioral
Decision Making, 12, 1999, pp. 37-53
22. Heerde, H. J. Van, Leeflang, P. S. H. and Wittink, D. R. Flexible decomposition of price
promotion effects using store-level scanner data, 2002b
23. Heerde, H. J. Van, Leeflang, P. S. H. and Wittink, D. R. How promotions work: Scan*Pro-based
evolutionary model building, Schmalenbach Business Review, 54, 2002a, pp. 198-
220
24. Hendry, D. F. Econometric modeling, Lecture notes for the PhD course in econometric modelling
and economic forecasting, Department of Economics, University of Oslo, 2000
25. Holland, J. H. Adaptation in natural and artificial systems, University of Michigan Press, Ann
Arbor, MI, USA. (extended new edition, MIT Press, Cambridge, MA, USA, 1992), 1975
26. Klassen, R. D., and Flores, B. E. Forecasting practices of Canadian firms: Survey results and
comparisons, International Journal of Production Economics, 70, 2001, pp. 163-174

480
 Quantitative Methods Inquires

27. Kleinmutz, B. Why we still use our heads instead of formulas: towards an integrative
approach, Psychologieal Bulletin, 107, 1990, pp. 296-310
28. Lilliefors, H. W. On the Kolmogorov-Smirnov test for normality with mean and variance
unknown. Journal of the American Statistical Association, 64, 1967, pp. 399-402
29. Lim, J. S. and O'Connor, M. Judgmental forecasting with time series and causal
information, International Journal of Forecasting, 12, 1996, pp. 139-153
30. Liu, Z., Zhou, J., and Lai, S. New adaptive genetic algorithm based on ranking, Proceedings
of the Second International Conference on Machine Learning and Cybernetics, 2003
31. Makridakis, S. and Maclachlan, D. Judgmental biases in applied decision making situations,
Working Paper, INSEAD, Fountainbleau, France, 1984
32. Mathews, B. P. and Diamantopoulos, A. Managerial intervention in forecasting: An empirical
investigation of forecast manipulation, International Journal of Research in
Marketing, 3, 1986, pp. 3-10
33. Nikolopoulos, K., Fildes, R., Goodwin, P. and Lawrence, M. On the accuracy of judgemental
interventionson forecasting support systems, Working paper 2005/022, Lancaster
University Management School, 2005
34. Pham, D. T. and Karaboga, D. Genetic algorithms with variable mutation rates: Application
to fuzzy logic controller design, Proceedings of the I MECH E Part I. Journal of
Systems & Control Engineering, 211(2), 1997, pp.157–167
35. Rawlings, J. O., Pantula, S. G. and Dickey, D. A. Applied regression Analysis—A Research
Tool, SpringerVerlag, New York INC. New York, NY, USA, 1998, pp. 371-372
36. Remus, W., O’Connor, M. and Griggs, K. Does reliable information improve the accuracy of
judgmental forecasts? International Journal of Forecasting, 11, 1995, pp. 285-293
37. Salo, A. A. and Bunn, D. W. Decomposition in the assessment of judgmental probability
forecasts, Technological forecasting and Social Change, 149, 1995, pp. 13-25
38. Sanders, N. and Ritzman, L. P. The need for contextual and technical knowledge in
judgmental forecasting, Journal of Behavioral Decision Making, 5, 1992, pp. 39-52
39. Sanders, N. R. and Manrodt, K. B. Forecasting practices in US corporations: Survey results,
Interfaces 24, 1994, pp. 92-100
40. Scapolo, F. and Miles, I. Eliciting experts’ knowledge: A comparison of two methods,
Technological Forecasting & Social Change, 73, 2006, pp. 679-704
41. Schaffer, J. D., Caruana, R. A., Eshelman, L. J., and Das, R. A study of control parameters
affecting online performance of genetic algorithms for function optimization, In
Proceedings of the Third International Conference on Genetic Algorithms, 1989, pp. 51–
60
42. Skalak, D. B. Prototype and feature selection by sampling and random mutation hill
climbing algorithms, In “Proceedings of the Eleventh International Conference on
Machine Learning”, New Brunswick, NJ. Morgan Kaufmann, 1994, pp. 293-301
43. Syntetos, A. A., Nikolopoulos, K., Boylan, J. E., Fildes, R. and Goodwin, P. The effects of
integreating management judgement into intermittent demand forecasts,
International Journal of Production Economics, 118, 2009, pp. 72-81
44. Tversky, A. and Kahneman, D. Judgement under uncertainty: Heuristics and biases, Science,
185, 1974, pp. 1124-1131
45. Wang, C. L., and Wang, L. C. A GA-based sales forecasting model incorporating promotion
factors, Working Paper., Department of Industrial Engineering and Enterprise
information, Tunghai University, Taiwan, ROC, 2009
46. Williams, T. M. Adaptive Holt-Winters forecasting, Journal of Operational Research Society, 38
(6), 1987, pp. 553-560

481
 Quantitative Methods Inquires

Appendix A.

Table A1. Results of model checking

Sample A Sample B
a
item Normality CI (mean of max) ACF Normality CI (mean of max) ACF
b c
1 O 105.341 X O 93.677 X
2 O 118.517 X O 6.358 X
3 O 116.247 O O 82.962 O
4 O 51.337 O O 78.099 O
5 O 67.519 O O 170.917 X
6 O 21.686 O O 113.379 O
7 O 77.332 O O 91.224 X
8 O 40.015 O O 118.214 O
9 O 195.611 O O 227.116 O
10 O 374.608 O O 191.684 X
Note: a. CI denotes condition index.
b. O denotes a success to pass the test.
c. X denotes a failure to pass the test.

1
Chin-Lien Wang, is a PhD candidate of Industrial Engineering & Enterprise Information at Tunghai University in
Taiwan. He got his MBA from Michigan State University in 1985, and got his MS in computer science from DePaul
University in 1989. At present, he teaches and does research at the Department of Business Administration, Ling
Tung University in Taiwan. His current research interest includes business modeling, search methods, and time
series forecasting.
Department of Industrial Engineering and Enterprise information, Tunghai University, 181 Section 3, Taichung
Harbor Rd. Taichung, 40704 Taiwan, ROC
Department of Business Administration, Ling Tung University, 1 Ling Tung Rd. Taichung,
Taiwan, ROC

2
Dr. Li-Chih Wang is a Professor and Chairman of Industrial Engineering and Enterprise Information at Tunghai
University, Taiwan. He earned his B.S. degree in Industrial Engineering from Tunghai University, Taiwan and M.S.
and Ph.D. degrees in Industrial and Systems Engineering from The Ohio State University, U.S.A. His teaching covers
logistics and supply chain management, systems analysis, manufacturing planning and control system, and ERP-II at
the levels of undergraduate, master's, and doctorate.
Dr. Wang is currently working on supply chain management (SCM) and advanced planning & scheduling (APS)
system development. He has conducted extensive research leading to the current academic and industrial projects
(e.g., memory module, electronic assembly, TFT-LCD). Dr. Wang has published three SCM/APS related books which
are widely referenced in Taiwan and numerous scholarly papers and reports in these areas. His research has been
supported by both government agencies (e.g., National Science Council, Ministry of Education, Ministry of Economy)
and industry (e.g., Kingston, Macronix, EverTek, Farcent).
Dr. Li-Chih Wang has consulted for a number of high tech industries and served as Chief Advisor for developing
Digichain’s supply chain planning (SCP) and APS software/system which has been implemented in a number of high
tech industries. He also actively promotes SCM concept and system in government, research institute, academy and
industry for the last 15 years.

3
Codification of references:
Armstrong, S. Forecasting by extrapolation: Conclusion from 25 years of research, Interfaces 14, 6,
[1]
1984, pp. 52-66
Dalrymple, D. J. Sales forecasting practices, results from a United States survey, International Journal
[2]
of Forecasting, 3, 1987, pp. 379-391
Kleinmutz, B. Why we still use our heads instead of formulas: towards an integrative approach,
[3]
Psychologieal Bulletin, 107, 1990, pp. 296-310
Bunn, D. W. and Salo, A. A. Adjustment of forecasts with model consistent expectations, International
[4]
Journal of Forecasting, 12, 1996, pp. 163-170
Sanders, N. and Ritzman, L. P. The need for contextual and technical knowledge in judgmental
[5]
forecasting, Journal of Behavioral Decision Making, 5, 1992, pp. 39-52
Makridakis, S. and Maclachlan, D. Judgmental biases in applied decision making situations, Working
[6]
Paper, INSEAD, Fountainbleau, France, 1984

482
 Quantitative Methods Inquires

Mathews, B. P. and Diamantopoulos, A. Managerial intervention in forecasting: An empirical

[7] investigation of forecast manipulation, International Journal of Research in Marketing, 3,
1986, pp. 3-10
Klassen, R. D., and Flores, B. E. Forecasting practices of Canadian firms: Survey results and
[8]
comparisons, International Journal of Production Economics, 70, 2001, pp. 163-174
Goodwin, P. and Fildes, R. Judgmental forecasts of time series affected by special events: Does
[9] providing a statistical forecast improve accuracy? Journal of Behavioral Decision Making,
12, 1999, pp. 37-53
Blattberg. R. C. and Hoch, S. J. Database models and managerial intuition! 50% model + 50%
[10]
manager, Management Science, 36. 8, 1990, pp. 887-899
Nikolopoulos, K., Fildes, R., Goodwin, P. and Lawrence, M. On the accuracy of judgemental
[11] interventionson forecasting support systems, Working paper 2005/022, Lancaster
University Management School, 2005
Fildes, R. and Goodwin,P. Good and bad judgement in forecasting: Lessons from four companies,
[12]
Foresight, Fall, 2007, pp. 5-10
Remus, W., O’Connor, M. and Griggs, K. Does reliable information improve the accuracy of
[13]
judgmental forecasts? International Journal of Forecasting, 11, 1995, pp. 285-293
Lim, J. S. and O'Connor, M. Judgmental forecasting with time series and causal information,
[14]
International Journal of Forecasting, 12, 1996, pp. 139-153
Scapolo, F. and Miles, I. Eliciting experts’ knowledge: A comparison of two methods, Technological
[15]
Forecasting & Social Change, 73, 2006, pp. 679-704
Armstrong, S., Collopy, F. and Thomas, Y. J. Decomposition by causal forces: a procedure for
[16]
forecasting complex time series, International Journal of Forecasting, 21, 2005, pp. 25-36
Edmundson, R. H., Decomposition: A strategy for judgmental forecasting, Journal of Forecasting, 19,
[17]
pp. 305-314, IEEE Transactions on Evolutionary Computation, 3(2), 1990, pp. 124-141
Salo, A. A. and Bunn, D. W. Decomposition in the assessment of judgmental probability forecasts,
[18]
Technological forecasting and Social Change, 149, 1995, pp. 13-25
Tversky, A. and Kahneman, D. Judgement under uncertainty: Heuristics and biases, Science, 185,
[19]
1974, pp. 1124-1131
[20] Armstrong, S. Long-range forecasting: from crystal ball to computer, New York; Wiley, 1985
Sanders, N. R. and Manrodt, K. B. Forecasting practices in US corporations: Survey results, Interfaces
[21]
24, 1994, pp. 92-100
Box, G. E. P., Jenkins, G. M., and Reinsel, G. C. Time series analysis: Forecasting and control (3rd ed.),
[22]
Englewood Cliffs, NJ7 Prentice Hall, 1994
De Gooijer, J. G. and Hyndman, R. J. 25 years of time series forecasting, International Journal of
[23]
Forecasting, 22, 2006, pp. 443- 473
Foekens, E. W., Leeßang, P. S. H. and Wittink, D. R.. Varying parameter models to accommodate
[24]
dynamic promotion effects, Journal of Econometrics, 89, 1999, pp. 249-268
Heerde, H. J. Van, Leeflang, P. S. H. and Wittink, D. R. How promotions work: Scan*Pro-based
[25]
evolutionary model building, Schmalenbach Business Review, 54, 2002a, pp. 198-220
Heerde, H. J. Van, Leeflang, P. S. H. and Wittink, D. R. Flexible decomposition of price promotion effects
[26]
using store-level scanner data, 2002b
Carroll. R. J. and Ruppert. D. Transformation and Weighting in Regression, Chapman and Hall, New
[27]
York, NY, USA, 1998, pp. 115-160
Franses, P. H. and McAleer, M. Testing for unit roots and non-linear transformations, Journal of Times
[28]
Series Analysis, 19, 1998, pp. 147-164
Belsley, D.A. Assessing the presence of harmful collinearity and other forms of weak data through a
[29]
test for signal-to-noise, Journal of Econometrics, 20, 1982, pp. 211-253
Belsley, D. A., Kuh, E., and Welsch, R. E. Regression Diagnostics. Identifying Influential Data and
[30]
Sources of Collinearity, John Wiley & Sons, New York, 1980
Hendry, D. F. Econometric modeling, Lecture notes for the PhD course in econometric modelling and
[31]
economic forecasting, Department of Economics, University of Oslo, 2000
Wang, C. L., and Wang, L. C. A GA-based sales forecasting model incorporating promotion factors,
[32] Working Paper., Department of Industrial Engineering and Enterprise information, Tunghai
University, Taiwan, ROC, 2009
Holland, J. H. Adaptation in natural and artificial systems, University of Michigan Press, Ann Arbor, MI,
[33]
USA. (extended new edition, MIT Press, Cambridge, MA, USA, 1992), 1975
Goldberg, D. E. Simple genetic algorithms and the minimal deceptive problem, in: Davis, L. (ed.)
[34] “Genetic Algorithms and Simulated Annealing”, Hyperion Books, New York, NY, USA, Morgan
Kaufmann, San Francisco, CA, USA, 1987, pp. 74-88
Goldberg, D. E. Genetic Algorithms, in Search, Optimization & Machine Learning, Addison-Wesley,
[35]
Boston, MA, USA, 1989

483
 Quantitative Methods Inquires

De Jong, K. A. and Spears, W. M. A formal analysis of the role of multi-point crossover in genetic
[36]
algorithms, Annals of Mathematics and Artificial Intelligence 5(1), 1992, pp. 1–26
Skalak, D. B. Prototype and feature selection by sampling and random mutation hill climbing
[37] algorithms, In “Proceedings of the Eleventh International Conference on Machine Learning”,
New Brunswick, NJ. Morgan Kaufmann, 1994, pp. 293-301
Liu, Z., Zhou, J., and Lai, S. New adaptive genetic algorithm based on ranking, Proceedings of the
[38]
Second International Conference on Machine Learning and Cybernetics, 2003
Pham, D. T. and Karaboga, D. Genetic algorithms with variable mutation rates: Application to fuzzy
[39] logic controller design, Proceedings of the I MECH E Part I. Journal of Systems & Control
Engineering, 211(2), 1997, pp.157–167
Schaffer, J. D., Caruana, R. A., Eshelman, L. J., and Das, R. A study of control parameters affecting
[40] online performance of genetic algorithms for function optimization, In Proceedings of the
Third International Conference on Genetic Algorithms, 1989, pp. 51–60
Eiben, A. E., Hinterding, R. and Michalewicz, Z. Parameter control in evolutionary algorithms, IEEE
[41]
Transactions on Evolutionary Computation, 3(2), 1999, pp. 124-141
Lilliefors, H. W. On the Kolmogorov-Smirnov test for normality with mean and variance unknown.
[42]
Journal of the American Statistical Association, 64, 1967, pp. 399-402
Berilant, J., de wet, T., and Goegebeur, Y. A goodness-of-fit statistic for Pareto-type behavior. Journal
[43]
of Computational and Applied Mathematics, 186, 1, 2005, pp. 99-116
Rawlings, J. O., Pantula, S. G. and Dickey, D. A. Applied regression Analysis—A Research Tool,
[44]
SpringerVerlag, New York INC. New York, NY, USA, 1998, pp. 371-372
Williams, T. M. Adaptive Holt-Winters forecasting, Journal of Operational Research Society, 38 (6), 1987,
[45]
pp. 553-560
Syntetos, A. A., Nikolopoulos, K., Boylan, J. E., Fildes, R. and Goodwin, P. The effects of integreating
[46] management judgement into intermittent demand forecasts, International Journal of
Production Economics, 118, 2009, pp. 72-81

484
 Quantitative Methods Inquires

INDUCTION OF MEAN OUTPUT PREDICTION TREES

FROM CONTINUOUS TEMPORAL METEOROLOGICAL DATA1

Dima ALBERG2
PhD Candidate, Department of Information Systems Engineering,
Ben-Gurion University of the Negev, Beer-Sheva, Israel

E-mail: [email protected]

Mark LAST3
PhD, Associate Professor, Department of Information Systems Engineering,
Ben-Gurion University of the Negev, Beer-Sheva, Israel

E-mail: [email protected]

Avner BEN-YAIR4,5
PhD, Department of Industrial Engineering and Management,
Sami Shamoon College of Engineering, Beer-Sheva, Israel

E-mail: [email protected]

Abstract: In this paper, we present a novel method for fast data-driven construction of
regression trees from temporal datasets including continuous data streams. The proposed
Mean Output Prediction Tree (MOPT) algorithm transforms continuous temporal data into two
statistical moments according to a user-specified time resolution and builds a regression tree
for estimating the prediction interval of the output (dependent) variable. Results on two
benchmark data sets show that the MOPT algorithm produces more accurate and easily
interpretable prediction models than other state-of-the-art regression tree methods.

Key words: temporal prediction; inductive learning; time resolution; regression trees; split
criteria; multivariate statistics; multivariate time series

1. Introduction

The time dimension is one of the most important attributes of massive continuous
temporal data sets or continuous data streams [13]6, where data arrives and has to be
processed on a continuous basis. Sources of such data include real-time monitoring devices
as: meteorological stations, traffic control systems, financial markets, etc. If we extract a
portion of data arrived over a finite time period and store it in a persistent database, it
becomes a temporal dataset. Generally, the time dimension is represented in a temporal

485
 Quantitative Methods Inquires

dataset as a calendar variable which has an agglomerative structure consisting of several

time granules [22]: for example, 60 seconds represent 1 minute, 60 minutes represent one
hour, etc. The correct choice of the pre-processing time granularity very often predetermines
the accuracy and interpretability of data stream mining algorithms.
We are interested in prediction of temporal continuous variables, since they are
abundant in most data streams mentioned above. However, the existing prediction methods
(such as Regression Tree Models [3, 6, 18-19, 21, 24-25]) do not take the time dimension
and time granularity into account, since they were developed for mainly static (time-
invariant) databases. In this work, we present a new prediction algorithm capable to induce
an accurate and interpretable model for estimating the values of a given continuous output
variable in a massive temporal data set or a continuous data stream.

2. Problem Statement and Prior Work

In many data streams, the data is available as time-continuous statistical moments

(mean, variance, etc.) calculated over pre-defined measurement cycles rather than raw
values sampled at discrete points in time. Examples of such data streams include:
meteorological data, financial data, factory control systems, sensor networks, etc. For
example, a meteorological station might be continuously storing mean and variance
estimation for a large number of meteorological attributes at predefined time intervals (e.g.,
every 10 minutes). A prediction model that is built using multiple statistical moments instead
of discretely sampled exact values is likely to have a lower update cost, since as long as an
attribute value remains within the prediction interval, fewer updates to the model will be
required.
However, supervised predictive data mining models like regression models (GLM,
MARS[10]) and Regression Trees (M5 [21], M5' [25], CART [3], GUIDE [19], RETIS [18],
MAUVE [24], (M)SMOTI [6]) widely used at present for prediction of continuous target
variables do not utilize multiple statistical moments of input and target attributes. Time-
series prediction models (ARIMA, ARCH [9], and GARCH [2]), which carry out simultaneous
prediction of continuous target variables represented by statistical moments are frequently
non stable on significant volumes of non-stationary data and require labor-consuming
reassessment at uncertain time intervals [1]. Another difficulty is to produce an interpretable
set of prediction rules for such cases. Indeed, even supposing that it would be possible to
build an accurate regression tree or a set of logical rules using the time dependent input
attributes, the resulting model is likely to be very intricate and essentially impossible to
interpret [12].
Interval prediction is an important part of the forecasting process aimed at
enhancing the limited accuracy of point estimation. An interval forecast usually consists of an
upper and a lower limit between which the future value is expected to lie with a prescribed
probability. The limits are sometimes called forecast limits [26] or prediction bounds [5],
whereas the interval is sometimes called a confidence interval [12] or a forecast region [16].
We prefer the more widely-used term prediction interval, as used by Chatfield [7] and
Harvey [14], both because it is more descriptive and because the term confidence interval is
usually applied to interval estimates for fixed but unknown parameters. In our case, a
prediction interval is an interval estimate for an (unknown) future value of the output
(dependent) variable. Since a future value can be regarded as a random variable at the time

486
 Quantitative Methods Inquires

the forecast is made, a prediction interval involves a different sort of probability statement
from that implied by a confidence interval. The above considerations cause a need of a
model which can process the incoming data in real time and on the basis of the received
results to make interval prediction of target time dependent continuous variables with a
given level of statistical confidence.

2. The Proposed Data Stream Mining Methodology

2.1. The Model Induction Algorithm Overview

The proposed algorithm is aimed at inducing a prediction model for a case where
every input (predictive) temporal variable is continuous and in a given sliding window k , the
two statistics of mean and variance are calculated for each measurement cycle. The output
(predicted) variable is the mean value of a continuous temporal variable calculated for the
future sliding window k + Δ .
As inputs the algorithm receives the time resolution interval j , the two first
statistical moments of each temporally continuous input variableX with user predefined lag
Δ history and temporally continuous output variable Y , as well as the significance level α .
In the conventional regression tree algorithm, the objective is to build an inductive predictor
assuming the following functional form:

Y = {yˆ jk } = f {x jk − Δ } (1)

where the predicted target variable in a sliding window k is represented as a function of

input numerical variables in k − Δ sliding window, where Δ is a user-specified prediction
lag parameter.
The proposed algorithm will build an inductive predictor of the following form:

{ ( ( ))} = f {x
Y = yˆ jk , w T ΘY jk jk − Δ , sˆx2jk −Δ }, (2)

where

( ) (
T Θ Y jk = w Θ Y jk − Θ Y j ) Ψ(Θ)(Θ
T
Y jk ) { }
− Θ Yl , Θ Y jk ∈ y jk , sˆ y2 jk . (3)

Where for time resolution j in sliding window k , ŷ jk is the predicted value of

temporally continuous output variable Y , Θ Y jk is the mixture mean variance parameter for

output variable ( ( )) is the mixture density estimation weight for the output
Y , w T Θ Y jk
2
variable Y and x jk − Δ , sˆ x jk − Δ are mean and standard deviation estimators of a temporally

continuous input variable X for time resolution j in sliding window k − Δ . Finally, in (3)
for time resolution j in the sliding window k , the joint variable T Θ Y jk ( ) is the mean-

487
 Quantitative Methods Inquires

variance estimator based on the two first statistical moments of the output variable Y , ΘY j
is the vector of the means of the parameter Θ Y jk and Ψ (Θ ) represents the within-group
normalized covariance matrix for parameter Θ Y jk , which estimates normalized mean

variance covariance values of the predicted output variable Y . The confidence interval of the
( ) can be approximated using the F distribution as follows:
joint variable T Θ Y jk

( ( )) 2( −( 1)(− 2+) 1) ⋅ F ((α 2), (1 − α 2),2,

UB, LB T Θ Y jk =
j j
j − 2) , (4)
j j

where j is the number of measurement cycles in the sliding window k for time resolution
j and α is the user-specified significance level (default value is 0.05).
When the variance is independent of the mean value of the variable, the values of
( ) are expected to lie inside the confidence interval (see (4)) implying
the joint variable T Θ Y jk
that the interaction between mean and dispersion variables does not add information about
the behavior of the corresponding input variable. However, when some values of the joint
variable are found outside the boundaries UB and LB (see (4)), it can be said that the
interaction between mean and dispersion variables adds further information about the input
variable X . These outliers provide sufficient information for the output variable prediction
and therefore we can consider only the outliers when evaluating the candidate split points of
an input variable. In case when no outliers are found, the algorithm checks the possibility to
switch to the higher time resolution and if the higher resolution represents the initial (raw)
time resolution, the algorithm proceeds as the regular RETIS [18] algorithm.
The impurity merit (5.1) and (5.2) contains two parts, left and right, whereas the
major objective is to find the optimal split point X , which minimizes the expression in (6):

( ) ∑ w (T (Θ) ( ))
NL
Var T (Θ ) = p
L 2
−T Θ
L L L L
X i (5.1)
i =1

⎛
( ) ( ( ) ) ⎞⎟⎟
NR
⎜⎜Var T (Θ )R = p R ∑ w XR T (Θ )iR − T Θ
R 2
(5.2)
⎝ i =1 ⎠
( (
X * = arg min Var T (Θ ) + Var T (Θ )
T (Θ )
L
) ( R
)) (6)

T (Θ ) and T (Θ ) are the left (right) joint mean variance estimator values of target
L R
Here

variable Y , p L and p R are the relative number of N L ( N R ) cases that are assigned to the
left (right) child, while w XL ( w XR ) is the left (right) mixture density estimation weight for the
target variable Y .
Thus, the best split at a node is defined as the split, which minimizes the weighted
variance of the joint mean variance estimator.

488
 Quantitative Methods Inquires

2. 2 Performance Metrics
A validation dataset for time resolution j and sliding window w j is made up of
k ∈ {1,..., N } instances (sliding windows), each mapping an input vector ( x1 ,..., x A ) to a
k

given target y k . The error is given by: ek = yˆ k − y k , where ŷ k represents the predicted
value for the k -th input pattern (2). The overall performance is computed by a global metric,
namely the Mean Absolute Error (MAE) and Root Mean Squared (RMSE). However, the RMSE
is more sensitive to high volatility errors than MAE. In order to compare the accuracy of trees
from different domains the % of Explained Variability (EV) defined as:

EV (T ) =
(SSE (Mean) − SSE (T )) ⋅ 100%
(7)
SSE (Mean)

Here Mean is a majority rule predictor, which always predicts the mean value of the
training set, SSE (Mean ) and SSE (T ) are sum of square errors from the mean value and
the value predicted by the evaluated regression tree (T ) model, respectively. Another
possibility to compare regression tree models is the Cost Complexity Measure (CCM) defined
as:

CCM (T ) = RMSE (T ) + α ⋅ TS (T ) . (8)

Here RMSE (T ) is the estimated error cost of regression tree T , TS (T ) is the number of
terminal nodes in the tree, and α is the user defined non-negative cost complexity
parameter adopted from [9], where it is shown that for a given complexity parameter, there
is a unique smallest subtree of the saturated tree that minimizes the cost-complexity
measure, which actually quantifies the tradeoff between the size of the tree and how well
the tree fits the data.

3. Experimental Results

The performance of the MOPT algorithm proposed in the Section 2.1 was evaluated
on ElNino data set from the UCI Machine Learning Repository [23]. The selected data set
consist from numerical attribute types and belong to the multivariate spatio temporal
regression domain. Finally, the performance of the complete Mean Output Prediction Tree
(MOPT) algorithm was evaluated on the second data set, which represents a multivariate
continuous data stream collected at a meteorological station in Israel during a period of
about 8 years.
The algorithm performance is compared to four state-of-the-art prediction
algorithms implemented by the Java API of WEKA [25]: M5P Tree [25] (Bagging M5P tree),
M5-Rules [21] (Bagging M5-Rules), RepTree (Bagging RepTree) and by our implementation of
the RETIS [18] algorithm (RETIS-M). The main difference between RETIS[18] and RETIS-M
algorithm concludes in more fast splitting criterion implementation.

489
 Quantitative Methods Inquires

3.1. El Nino Data Set

The El Nino/Southern Oscillation (ENSO) cycle of 1982-1983, the strongest of the
century, created many problems throughout the world. The El Nino dataset consists of the
following attributes: buoy, date, latitude, longitude, zonal winds (west<0, east>0), meridian
winds (south<0, north>0), relative humidity, air temperature and sea surface temperature.
Data was taken from the buoys from as early as 1980 for some locations publicly available
UCI Machine Learning Repository [23]. Other data that was taken in various locations are
rainfall, solar radiation, current levels, and subsurface temperatures. The experimental data
is represented using a single (daily) time resolution and it consists of 178,080 data instances.
Important to note, that all data readings were taken at the same time of day and the target
(predicted) variable is the subsurface temperature.
Finally in order to evaluate the predictive performance, the set of all examples was
split into learning and testing examples sets in proportion 70:30.
The results in Table 1 show that under RMSE and Explained Variability criterions the
MOPT and the RETIS-M algorithms are more accurate than other proposed algorithms in
terms of t-test pair-wise difference. We have denoted by * the cases where the p value of the
difference between MOPT and other algorithms is smaller than or equal to 5%. The MOPT
algorithm outperforms significantly the other algorithms in the terms of cost complexity
measure. Finally, we will to consider that our proposed MOPT Tree is more interpretable
than RETIS-M tree in terms of Tree Size measure (7 vs. 23).

Table 1. El Nino data set learners comparison

Learner RMSE TS CCM EV
B-M5 Rules 0.84* 7 1.01* 0.46*
B-M5P Tree 0.83* 10 1.07* 0.47*
B-REPTree 1.57* 5 1.69* NA
M5 Rules 0.86* 7 1.03* 0.45*
M5P Tree 0.84* 8 1.03* 0.46*
MOPT 0.60 7 0.77 0.62
REPTree 1.57* 3 1.64* NA
RETIS-M 0.63 23 1.18* 0.60

3.2. Israel Meteorology Data Set

In this experiment we used the data collected at a meteorological station in Israel
during a period of about 8 years (from 01/07/1997 to 31/08/2005). Spatio- temporal
meteorological attributes (such as pressure, temperature, solar radiation, horizontal wind:
direction, speed, gust speed, gust time, and vertical wind: down-up and up direction) are
measured constantly in time and saved every 10 minutes in the form of mean and variance.
The selected data set exceeds 1,500,000 records. The total number of temporal and
meteorological attributes collected at the three stations is 22. Our first experiment was run
on the summer months (JUN, JUL and AUG) only. The experimental data was represented
using 5 time resolutions (10, 30, 60, 90 and 120 Minutes). The algorithms were run for
11:00-12:00 and 23:00 – 24:00 hours prediction.
The aim of this experiment was to compare the different state-of-the-art algorithms
for different time resolutions in order to be able to predict wind directions for short time
range (now-casting) up to 8 hours sliding window ahead. We have shown that the most
state-of-the-art algorithms gave the same or poorer quality of results and less interpretable

490
 Quantitative Methods Inquires

trees as the proposed Mean Output Prediction Tree (MOPT) algorithm. The results also
pinpoint the fact that sometimes there was no need to use very high time resolution, but only
lower time resolutions, since the statistical measures checked (i.e. RMSE, Cost Complexity
Measure and Percentage of Explained Variability) were similar.
Tables 2, 3 compare between the proposed MOPT algorithm and four state-of-the-
art algorithms: Modified RETIS (RETIS-M), M5P and REPTree in five time resolution scales in
terms of Cost Complexity and Explained Variability measures. For more effective evaluation
of the MOPT algorithm we performed short term prediction of 11:00 and 23:00 hour. The
sliding window size and the prediction lag were set to 8 hours and 3 hours, respectively.
Thus for predicting 11 hour wind direction we collected data from 00:00 to 8:00 and for
predicting 23 hour wind direction we collected data from 12:00 to 20:00. In each prediction
case, the main issue is to predict wind direction 3 hours ahead therefore fast robust and
accurate prediction algorithm producing a compact model is needed. For each time
resolution, we preprocessed the raw data and calculated the first two statistical moments for
each attribute in every measurement cycle. The MOPT algorithm refers to each input
attribute as a 2-dimensional array (two moments × number of instances) and determines the
split point with the aid of two moments target variable impurity and the variance of the input
variable. As in the previous experiments, the differences are considered statistically
significant when the p-value of the t-pair-wise test statistic is smaller than or equal to 5%
which signed by *.

Table 2. MOPT and state-of-the-arts models cost complexity Measure (CCM) cross
resolutions results for 11:00 hour prediction
TR MOPT RETIS-M M5P REPTree
10 116.61 227.28* 117.18 124.23
30 117.20 245.48* 149.48* 132.07
60 120.58 251.13* 172.78* 143.31
90 120.32 236.58* 171.03* 139.11
120 114.61 240.73* 188.88* 148.70*

In 10 minutes resolution the M5P slightly outperforms the proposed MOPT model
and significantly better than other models. In other resolutions the MOPT model significantly
better than other state-of-arts models. This result pinpoint to the fact that adding second
moments to split criterion improves quality of prediction for higher time resolution.

Table 3. MOPT and state-of-the-arts models cost complexity measure (CCM) cross
resolutions results for 23:00 hour prediction
TR MOPT RETIS-M M5P REPTree
10 43.95 60.04* 55.00 53.01
30 43.27 59.84* 60.53* 59.74*
60 45.45 62.28* 53.38 60.72*
90 44.55 59.76* 57.76 74.45*
120 44.53 61.82* 53.60 60.00*

By comparison to state-of-the-art algorithms, the MOPT algorithm demonstrates

more stable prediction accuracy with a more compact tree size in 23:00 hour prediction (for
example in 10 minutes resolution the size of MOPT tree is 502 versus 2947 of M5P). In this

491
 Quantitative Methods Inquires

case, the regression tree pruning procedure may significantly reduce the final size of the
tree, but this procedure is out of scope in the proposed MOPT approach because our main
purpose is to build accurate and compact tree with minimal access to the sliding window
training data.

Table 4. MOPT and state-of-the-arts models explained variability (%EV) cross resolutions
results for 11:00 hour prediction
TR MOPT RETIS-M M5P REPTree
10 45.24% 32.46% 60.18% 59.96%
30 45.35% 25.35% 41.99% 50.26%
60 44.59% 25.78% 29.73% 43.43%
90 44.32% 30.45% 29.69% 53.98%
120 48.65% 32.39% 23.09% 37.34%

Table 5. MOPT and state-of-the-arts models explained variability (%EV) cross resolutions
results for 23:00 hour prediction
TR MOPT RETIS-M M5P REPTree
10 29.3% -6.1% 22.7% 8.7%
30 30.7% -5.6% -10.4% 12.1%
60 25.6% -14.7% 8.5% 12.7%
90 28.5% -4.8% -13.3% -43.9%
120 28.3% -13.3% 12.9% 16.4%

The final stage of this experiment presented in the Tables 4-5 demonstrates the
comparison of Percentage of Explained Variability EV between five defined models and time
resolutions for 11:00 and 23:00 Hours respectively. The cells with negative explained
variability percent indicate the fact that the induced model is poorer (less accurate) than a
simple majority rule mean model. For example, RETIS-M, M5P and REPTree models have not
contributed to the explained variability of the 23:00 hour prediction. Important to
emphasize, that three state-of-the-art algorithms did not scale well to the low time
resolution of 120 minutes.

4. Conclusions

In this work, we have presented the two moments (mean-variance) Mean Output
Prediction Tree algorithm (MOPT), which is able to predict large amounts of massive
temporal data sets. The proposed algorithm differs from the state-of-the-art regression
algorithms in the splitting of each input and output feature to two moments according to the
input time resolution and it can also identify the most appropriate prediction time resolution
that minimizes the prediction error and builds more compact interval based regression tree.
The two conducted experiments indicate that the proposed algorithm produces
more accurate and compact models by comparison to the modern state-of-the-art regression
tree algorithms.

492
 Quantitative Methods Inquires

5. References

1. Alberg, D., Shalit, H. and Yosef, R. Estimating stock market volatility using asymmetric
GARCH models, Applied Financial Economics, vol. 18(15), 2008, pp. 1201-1208
2. Bollerslev, T.A. Conditionally heteroskedastic time series model for speculative prices and
rates of return, Review of Economics and Statistics, vol. 69, 1987, pp. 542–547
3. Breiman, L. and Friedman, J. Estimating Optimal Transformations for Multiple Regression
and Correlation, J. Amer. Statist. Assoc., vol. 80, 1985, p. 580
4. Breiman, L., Friedman, J., Olshen, R. and Stone, C. Classification and Regression Trees,
Wadsworth Int. Group, Belmont California USA, 1984
5. Brockwell, P. and Davis, R. Time Series: Theory and Methods, 2nd ed., New York: Springer-
Verlag, 1991
6. Ceci, M., Appice, A. and Malerba, D. Comparing simplification methods for model trees with
regression and splitting nodes, Foundations of Intelligent Systems, 14th International
Symposium, vol. 2871 of LNAI, 2003, pp. 49–56
7. Chatfield, C. The Analysis of Time Series, 5th ed., London: Chapman and Hall, 1995, pp. 431-
441
8. Cortez, F. and Morais, R.D. A data mining approach to predict forest fires using
meteorological data, in Neves, J., Santos, M.F. and Machado, J. (eds.) "New Trends in
Artificial Intelligence", Proceedings of the 13th EPIA 2007 - Portuguese Conference on
Artificial Intelligence, 2007, pp. 512-523
9. Engle, R. Autoregressive conditional heteroskedasticity with estimates of the variance of
United Kingdom inflation, Econometrica, vol. 50, 1982, pp. 987-1007
10. Friedman, J. Multivariate adaptative regression splines, Annals of Statistics, 1991, pp. 1-19
11. Gama, J., Rocha, R., Medas, P., Wah, B. and Wang, J. Accurate decision trees for mining high-
speed data streams, KDD 2003, 2003 , pp. 523-528
12. Granger, C. and Newbold, P. Forecasting Economic Time Series, 2nd ed., Academic Press,
New-York, 1986
13. Han, J., Cai, D. and Chen, Y. Multi dimensional analysis of data streams using stream
cubes in data streams models and algorithms, ed. Aggarwal, C., 2007, pp. 103-
125
14. Harvey, A. Forecasting Structural Time Series Models and the Kalman Filter, C U P,
Cambridge, 1989
15. Hulten, G., Spencer, L. and Domingos, P. Mining time-changing data stream, Proceedings of
the 7th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, 2001, pp. 97-106
16. Hyndman, R. Highest-density forecast regions for non-linear and non-normal time series
models, Journal of Forecasting, vol. 14, 1996
17. Kamber, M. and Han, J. Data Mining: Concepts And Techniques, 2nd ed., The Morgan
Kaufmann Series in Data Management Systems, 2007
18. Karalic, A. Linear regression in regression tree leaves, in “Proceedings of International School
for Synthesis of Expert”, vol. 10(3), 1992, pp. 151-163
19. Loh, W.Y. Regression tree models for designed experiments, 2nd Lehmann Symposium,
Knowledge, 2006, pp. 210-228
20. Ma, J., Zeng, D. and Chen, H. Spatial-temporal cross-correlation analysis: a new measure
and a case study in infectious disease, in Mehrotra, S. et. al. (eds ) Informatics, ISI,
LNCS 3975, 2006, pp. 542-547
21. Quinlan, J. Learning with continuous classes, in “Proceedings of the 5th Australian Joint
Conference on Artificial Intelligence”, Singapore World Scientific, 1992
22. Spranger, S. and Bry, F. Temporal data modeling and reasoning for information statistical
view of boosting, The Annals of Statistic, vol. 38(2), 2006, pp. 337-374

493
 Quantitative Methods Inquires

23. Vens, C. and Blockeel, H. A simple regression based heuristic for learning model trees,
Intelligent Data Analysis, vol. 10(3), 2006, pp. 215-236
24. Wang, Y. and Witten, I. Inducing of model trees for predicting continuous classes, in
“Proceedings of the 9th European Conference on Machine Learning”, Springer-Verlag,
1997, pp. 128-137
25. Wei, S., Dong, Y. and Pei, J.-B. Time Series Analysis, Redwood City Cal Addison-Wesley, 1990
26. * * * UCI, Machine Learning Repository, http://archive/ics/uci edu/ml/index html

1
Acknowledgment: This study was partially supported under a research contract from the Israel Ministry of
Defense.

2
Dima Alberg is currently Ph.D. candidate in the Department of Information Systems Engineering, Ben-Gurion
University of the Negev under the supervision of Prof. Mark Last. He is also Engineer in the Industrial Engineering
and Management Department of SCE - Shamoon College of Engineering. Dima was born in Vilnius, and repatriated
to Israel in 1996. He received his B.A and M.A. in Economics and Computer Science from Ben-Gurion University of
the Negev. His current research interests are in time series data mining, data streams segmentation and computer
simulation. His recent publications include the Journal of Business Economics and Management, Applied Financial
Economics, and Communications in Dependability and Quality Management.

3
Mark Last is currently Associate Professor at the Department of Information Systems Engineering, Ben-Gurion
University of the Negev, Israel and Head of the Software Engineering Program. Prior to that, he was a Visiting
Research Scholar at the US National Institute for Applied Computational Intelligence, Visiting Assistant Professor at
the Department of Computer Science and Engineering, University of South Florida, USA, Senior Consultant in
Industrial Engineering and Computing, and Head of the Production Control Department at AVX Israel. Mark
obtained his Ph.D. degree from Tel-Aviv University, Israel in 2000. He has published over 140 papers and chapters
in scientific journals, books, and refereed conferences. He is a co-author of two monographs and a co-editor of
seven edited volumes. His main research interests are focused on data mining, cross-lingual text mining, software
testing, and security informatics.
Prof. Last is a Senior Member of the IEEE Computer Society and a Professional Member of the Association for
Computing Machinery (ACM). He currently serves as Associate Editor of IEEE Transactions on Systems, Man, and
Cybernetics, where he has received the Best Associate Editor Award for 2006, and Pattern Analysis and Applications
(PAA). Prof. Last is very active in organizing cooperative international scientific activities. He has co-chaired four
international conferences and workshops on data mining and web intelligence.

4
Avner Ben-Yair is lecturer in the Industrial Engineering and Management Department, Sami Shamoon College of
Engineering, Israel. He was born in Moscow in 1961. He received his B.Sc. in Mechanical Engineering from the
Moscow Polygraphic Institute, Russia, and his M.Sc. degree in Health and Safety Engineering and Management
(Summa Cum Laude) from the Ben Gurion University of the Negev, Israel. He also received his Ph.D. degree in
Industrial Engineering and Management from the Ben Gurion University of the Negev, Israel. His professional
experience includes 13 years of engineering and management positions in Israeli chemical, pharmaceutical and
high-tech industries. His current research interests are in economic aspects of safety, reliability and failure analysis,
trade-off optimization models for organization systems, production planning, scheduling and control, cost
optimization and PERT-COST models, and strategic management. He has published 40 articles in various scientific
sources. His recent publications have appeared in Mathematics and Computers in Simulation, International Journal
of Production Economics, Communications in Dependability and Quality Management, and Computer Modelling
and New Technologies.

5
Corresponding author

6
Codification of references:
Alberg, D., Shalit, H. and Yosef, R. Estimating stock market volatility using asymmetric GARCH
[1]
models, Applied Financial Economics, vol. 18(15), 2008, pp. 1201-1208
Bollerslev, T.A. Conditionally heteroskedastic time series model for speculative prices and rates of
[2]
return, Review of Economics and Statistics, vol. 69, 1987, pp. 542–547
Breiman, L. and Friedman, J. Estimating Optimal Transformations for Multiple Regression and
[3]
Correlation, J. Amer. Statist. Assoc., vol. 80, 1985, p. 580
Breiman, L., Friedman, J., Olshen, R. and Stone, C. Classification and Regression Trees, Wadsworth Int.
[4]
Group, Belmont California USA, 1984
[5] Brockwell, P. and Davis, R. Time Series: Theory and Methods, 2nd ed., New York: Springer-Verlag, 1991
Ceci, M., Appice, A. and Malerba, D. Comparing simplification methods for model trees with
[6] regression and splitting nodes, Foundations of Intelligent Systems, 14th International
Symposium, vol. 2871 of LNAI, 2003, pp. 49–56

494
 Quantitative Methods Inquires

[7] Chatfield, C. The Analysis of Time Series, 5th ed., London: Chapman and Hall, 1995, pp. 431-441
Cortez, F. and Morais, R.D. A data mining approach to predict forest fires using meteorological
data, in Neves, J., Santos, M.F. and Machado, J. (eds.) "New Trends in Artificial Intelligence",
[8]
Proceedings of the 13th EPIA 2007 - Portuguese Conference on Artificial Intelligence, 2007,
pp. 512-523
Engle, R. Autoregressive conditional heteroskedasticity with estimates of the variance of United
[9]
Kingdom inflation, Econometrica, vol. 50, 1982, pp. 987-1007
[10] Friedman, J. Multivariate adaptative regression splines, Annals of Statistics, 1991, pp. 1-19
Gama, J., Rocha, R., Medas, P., Wah, B. and Wang, J. Accurate decision trees for mining high-speed
[11]
data streams, KDD 2003, 2003 , pp. 523-528
Granger, C. and Newbold, P. Forecasting Economic Time Series, 2nd ed., Academic Press, New-York,
[12]
1986
Han, J., Cai, D. and Chen, Y. Multi dimensional analysis of data streams using stream cubes in data
[13]
streams models and algorithms, ed. Aggarwal, C., 2007, pp. 103-125
[14] Harvey, A. Forecasting Structural Time Series Models and the Kalman Filter, C U P, Cambridge, 1989
Hulten, G., Spencer, L. and Domingos, P. Mining time-changing data stream, Proceedings of the 7th
[15] ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2001,
pp. 97-106
Hyndman, R. Highest-density forecast regions for non-linear and non-normal time series models,
[16]
Journal of Forecasting, vol. 14, 1996
Kamber, M. and Han, J. Data Mining: Concepts And Techniques, 2nd ed., The Morgan Kaufmann Series
[17]
in Data Management Systems, 2007
Karalic, A. Linear regression in regression tree leaves, in “Proceedings of International School for
[18]
Synthesis of Expert”, vol. 10(3), 1992, pp. 151-163
Loh, W.Y. Regression tree models for designed experiments, 2nd Lehmann Symposium, Knowledge,
[19]
2006, pp. 210-228
Ma, J., Zeng, D. and Chen, H. Spatial-temporal cross-correlation analysis: a new measure and a
[20] case study in infectious disease, in Mehrotra, S. et. al. (eds ) Informatics, ISI, LNCS 3975,
2006, pp. 542-547
Quinlan, J. Learning with continuous classes, in “Proceedings of the 5th Australian Joint Conference on
[21]
Artificial Intelligence”, Singapore World Scientific, 1992
Spranger, S. and Bry, F. Temporal data modeling and reasoning for information statistical view of
[22]
boosting, The Annals of Statistic, vol. 38(2), 2006, pp. 337-374
[23] * * * UCI, Machine Learning Repository, http://archive/ics/uci edu/ml/index html
Vens, C. and Blockeel, H. A simple regression based heuristic for learning model trees, Intelligent
[24]
Data Analysis, vol. 10(3), 2006, pp. 215-236
Wang, Y. and Witten, I. Inducing of model trees for predicting continuous classes, in “Proceedings of
[25]
the 9th European Conference on Machine Learning”, Springer-Verlag, 1997, pp. 128-137
[26] Wei, S., Dong, Y. and Pei, J.-B. Time Series Analysis, Redwood City Cal Addison-Wesley, 1990

495
 Quantitative Methods Inquires

THE IMPACT OF PACING MODE AND THE ANATOMICAL

POSITION OF PACING LEAD ON THE
INCIDENCE OF HEART FAILURE1

Oana STANCU (DINA)2

PhD, Associate Professor
University of Medicine and Pharmacy “Carol Davila”, Bucharest, Romania

E-mail: [email protected]

Iulia TEODORESCU (GABOR)3

Cardiologist MD, PhD, Researcher
Clinical Hospital “Sfantul Ioan”, Bucharest, Department of Internal Medicine and Cardiology

E-mail: [email protected]

Catalina Liliana ANDREI

PhD, Associate Professor
University of Medicine and Pharmacy”Carol Davila”, Bucharest, Romania

E-mail: [email protected]

Emanuel RADU4
Cardiologist MD, Researcher
Clinical Hospital “Sfantul Ioan”, Bucharest, Department of Angiography

E-mail: [email protected]

Octavian ZARA5
Cardiologist MD, Researcher
Clinical Hospital “Sfantul Ioan”, Bucharest, Department of Angiography

E-mail: [email protected]

Abstract: In Romania in the last decade, pacing is playing an increasingly important role in
the management of cardiac disease. If, at first, attention of the cardiologists and researchers
was focusing on the electrical rather than functional effects of pacing, the fact that pacing the
RV may initially improve cardiac function but may induce heart failure over time, has led to a
change in direction.
This study evaluates comparative the clinical outcome as well the incidence and predictors of
heart failure in 38 patients with VVIR pacing, VDDR and DDDR pacing implanted in “Sf. Ioan”
Hospital, Bucharest, over a period of 2 years. We also intended to evaluate the long-term
effects of alternative right ventricular pacing sites on LVEF.

Key words: right ventricular pacing; pacemaker syndrome; RV pacing sites; VVIR; VDDR;
DDDR pacing modes

496
 Quantitative Methods Inquires

1. Introduction

VVI pacing alone has shown a higher risk of sudden death compared with non-
paced patients with a similar degree of heart failure [1]6. A retrospective study of long-term
follow-up between VVI and DDD pacing showed that DDD enhances survival compared with
VVI in patients with heart failure and AV block [2]. Nielsen et al. demonstrated that VVI
pacing for sinus node disease was associated with the development of congestive heart
failure over a 5 year follow-up period.
On the other site the expectation that the hemodynamic benefits of atrioventricular
synchrony would lead to a reduction in cardiac mortality, a reduced risk of heart failure, and
a better quality of life were not proven by all the clinical trials. The MOST study which
followed for three years the cardiovascular mortality and morbidity in patients with DDDR
cardio stimulation toward patients with VVIR cardio stimulation showed no statistical
differences between the two groups. In exchange, concerning the heart failure episodes and
the quality of life, the study proved the superiority of the DDDR stimulation.
During ventricular pacing the asynchronous ventricular activation may lead to
abnormal regional myocardial blood flow and metabolic disturbancies which can reduce
systolic and diastolic left ventricular (LV) function[3,5]. These functional abnormalities seam
to have enhanced effects over time. Some studies have shown that long-term right
ventricular apical (RVA) pacing induces abnormal histologic changes and asymmetrical LV
hypertrophy and thinning [6,7,9]
The choice of pacing site in the right ventricle is an other important issue. No
recommendation could be made so far concerning the location of the right ventricular
pacing site.
The right ventricular apex, does not seem to lead to best haemodynamic results,
although it is easy accessible and best for electrode stability. [10,11,12]. While acute
haemodynamic studies find that outflow tract or dual-site pacing are best for haemodinamic
reasons, most of the controlled studies with permanent pacing found no significant
difference to right ventricular apical pacing
In a previous clinical study, on 547 patients we have shown that the relation
between VVIR pacing and the development of the pacemaker syndrome is likely to be
complex. Age, comorbidity and haemodinamic status before pacing are factors that influence
the appearance of the pacemaker syndrome. The patient group over 85 years had a higher
incidence of worsening heart failure than the other age groups. The patients with EF> 40%
before pacing had a better outcome than those with impaired left ventricular systolic
function.
The data of our previous study has shown that VVIR pacing may not induce directly
heart failure but may increase the risk of developping atrial fibrillation, an important
precipitant of heart failure.
In animal studies, pacing at the right ventricular outflow tract (RVOT) has been
shown to decrease the asynchrony of activation so that it seems to ameliorate the reduction
in LV function and prevent the wall motion abnormalities and the impair of the LV function.

497
 Quantitative Methods Inquires

2. Methods

The study, included 38 patients, men and women, which needed permanent
pacing, hospitalized in the Internal Medicine And Cardiology Department of “Sfantul Ioan”
Hospital, over a two and a half years period, between January 2007 and June 2009.
Patients who refused to sign the written consent and those with serious (severe) coagulation
disorders, chronic patients with dialysis or with cancer in terminal stages were excluded.

The patients were admitted for pacemaker implantation due to following diseases:

9% 28%
AF low rate
AV block III
63%
AV block II

Figure 1. Patients’ history

The follow up after the implant was planned at 1 month, 3 month, 12 month and 48
month.

PATIENT FOLLOW UP

50
40
30
20
10
0
1 MONTH 3 MONTHS 12 MONTHS 48 MONTHS

Figure 2. Patient follow up

We selected the pacing type, following a simple alghoritm that regarded the
aethiology and the anatomical and functional status of the atria (see figure 3).
All patients underwent implantation of a single chamber or a dual-chamber
pacemaker using one active fixation atrial lead and one active fixation bipolar ventricular
lead.
The ventricular pacing lead was inserted into the RV through subclavian vein
puncture. Under fluoroscopic guidance, the ventricular lead was positioned in the right
ventricular apex or in the RVOT ( by advancing the lead through the tricuspid valve and then
withdrawing it and positioning the tip against the interventricular septum and verifying the
position using multiple fluoroscopic views).

498
 Quantitative Methods Inquires

Paroxistic Bradicardia Permanent Bradicardia

VVIR at low heart Sinusal High grade AV block

rate in order not to Disfunction
interfere with atrial + AV
activity Conduction
Disturbance
AF Normal atria
or
dilated atria

Sinus Sinus node

rhythm Disfunction
DDDR

VVIR VDDR DDDR

Figure 3. Pacing’s type selection algorithm

The ventricular pacing leads were positioned at a stable position to obtain a

satisfactory pacing threshold value (mean 1.1 ± 0.2 V) at a pulse width of 0.5 ms and R-
wave sensing value (mean 12.3 ± 0.1 mV ). Atrial leads were positioned to the right atrial
lateral wall. After implantation, optimization of the AV delay was performed by using pulsed
Doppler echocardiography of the transmitral blood flow.
Patients were evaluated before implant by a complete clinical examination. Cardiac
risk factors, cardiac and associated non cardiac pathology were identified and concomitant
medication was recorded.
For a proper evaluation of heart failure a special attention was given to include the
patients in different NYHA classes according with their symptoms. The symptom screening,
prior to the clinical examination and echocardiogram was made by the physician by asking
the same questions in order to evaluate symptoms of heart failure.
The real effort capacity was estimated by standard 6 minutes walking test.
Before and after the implant , the end systolic and end diastolic volumes of the left
ventricle and the ejection fraction (Simpson method in two and four chambers incidence)
were measured.
Echocardiographic measurements were made in M mode and two-dimensional
echocardiography (2DE). Measurements of left ventricular end-diastolic volume (LVEDV), left
ventricular end-systolic volume (LVESV), and EF were obtained using the software installed on
the ultrasound equipment, with LVEDV measurements at the time of mitral valve closure and
LVESV measured on the image with the smallest LV cavity. The papillary muscles were
excluded from the volumes. Biplane Simpson's rule volumes were obtained from the apical
four- and two-chamber views.
M mode parameters were measured according to the American Society of
Cardiology.
The severity of MR was appreciated from Doppler color-flow in the conventional
parasternal long-axis and apical four-chamber images. Mitral regurgitation was

499
 Quantitative Methods Inquires

characterized as: mild (jet area/left atrial area <10%), moderate (jet area/left atrial area
10% to 20%), moderately severe (jet area/left atrial area 20% to 45%), and severe (jet
area/left atrial area >45%).

3. Results

As the result of the pre-procedural evaluation, the patients were included into three
groups with different pacing mode:

20

0 VDDR DDDR
VVIR
14 10 14

Figure 4. The distribution of the patients by the pacing mode

The incidence of heart failure at screening, at 3 months and at 12 months was the
following:

Table 1. The incidence of heart failure

Pacing mode Number of Number with Number with Number with

patients heart failure Heart failure heart failure
Screening At disclosure 12 month
VVIR 14 2 1 5
VDDR 10 3 0 2
DDDR 14 3 0 1
Total 38 8 1 8

At pre-discharge echocardiography, there was no significant difference from

baseline values in LVEF, cardiac output, as measured by Doppler echocardiography, or
transmitral A- and E-wave ratio in all the three groups.
Significant changes between the three pacing modes were found at the 3 months
and 12 months follow up, as shown in the table:

Table 2. Changes found at follow up checks

VVIR VDDR DDDR p value
NYHA class
ƒ Baseline 2.8 ± 0.3 2.5 ± 0.2 1.3 ± 0.2 < 0.01
ƒ 12 months 3.2 ± 0.3 2.4 ± 0.3 3.1 ± 0.3 < 0.05
QRS (ms )
ƒ Baseline 136 ± 22 130 ± 26 129 ± 27 < 0.01
ƒ 12 months 147 ± 26 132 ± 27 125 ± 18 < 0.01
LVEF (%)
ƒ Baseline 35 ± 6 31 ± 7 33 ± 6 < 0.01
ƒ 12 months 38 ± 5 30 ± 6 30 ± 7 < 0.05
Severe MR
ƒ Baseline 3 2 3 NS
ƒ 12 months 4 2 2 NS

500
 Quantitative Methods Inquires

Mitral regurgitation improved by at least one grade in 8 of 14 (57,14%) patients

with severe regurgitation.
The mean QRS duration was significantly longer in VVIR paced patients than in
VDDR or DDDR pacing.
We subdivided the patients in two different groups, those with right ventricular apex
pacing (RVA) and those with right ventricular outflow tract (RVOT ) pacing.

The differences between the 2 subgroups are seen in the following table:

Table 3. Subgroups’ particularities

Right ventricular RVOT/RS p
apex pacing pacing Value
Baseline echocardiography
LVEF 54 ± 11% 55 ±12% 0.72
E/A ratio 0.9 ± 0.3 1.1 ± 0.2 0.30
Optimal atrioventricular 140 ± 39 146 ± 36 0.47
interval(ms)
Discharge echocardiography
LVEF
E/A ratio 55 ±2% 58±11% 0.36
Optimal atrioventricular 1.0 ± 0.3% 1.1 ± 0.3% 0.25
interval (ms) 141 ± 38 145 ± 38 0.49
6 months echocardiography
LVEF 47 ±3% 56 ±12% 0.38
E/A ratio 0.9 ± 0.3% 1.0 ± 0.3% 0.28
Optimal atrioventricular 140 ± 38 146 ± 38 0.49
interval (ms)

4. Discussion

Although VVIR pacing is effective in preventing symptomatic bradyarrhythmias, it

has been demonstrated to be associated with a significant negative inotropic effect and with
an increased rate of congestive heart failure. This pacing modus has led also to a greater
increase in the grade of mitral regurgitation than the bicameral pacing modes.
RV apical pacing frequently produces an LBBB pattern with alteration in
myocardium depolarization and contraction which results in a reduction of the LVEF, an
prolongation of conduction intervals. Our study reveals that these changes are more
important over longer periods of time than immediately after pacing.
RVOT or RS pacing can improve cardiac performance, over that obtained with RV
apical pacing, despite the presence of AV synchrony. This improvement appears in every
pacing mode, but is more efficient in dual chamber pacing.

5. Conclusions

Even if for over 4 decades of cardiac pacing, the right ventricular apex (RVA) has
been the main site for right ventricular lead placement. during RVA pacing, larger QRS
duration, impairment of LV diastolic function with significant reduction in global LV function
is present.

501
 Quantitative Methods Inquires

Pacing at the RVOT is associated with more synchronous ventricular activation with
a narrower QRS duration and with lower incidence of deterioration in global LV systolic and
diastolic function.
Over long-term follow-up the clinical benefit seems to be greater.
RVOT pacing for routine pacemaker implantation might be the answer for
preventing and treating congestive heart failure in paced patients.
Further studies may be necessary in order to compare the benefits of RVOT pacing
compared with classical RVA pacing in patients with risk for heart failure.

References

1. Saxon, L.A., Stevenson, W.G., Middlekauff, H.R. and Stevenson L.W. Increased risk of
progressive haemodynamic deterioration in advanced heart failure patients
requiring permanent pacemakers, Am Heart J, 125, 1993, pp. 1306–1310
2. Alpert, M.A., Curtis, J.J., Sanfelippo, J.F. et al. Comparative survival after permanent
ventricular and dual chamber pacing for patients with chronic high degree
atrioventricular block with and without preexisting congestive heart failure, J
Am Coll Cardiol, 7, 1986, pp. 925–932
3. Rosenqvist, M., Isaaz, K., Botvinick, E.H., et al. Relative importance of activation sequence
compared to atrioventricular synchrony in left ventricular function, Am J Cardiol,
67, 1991, pp. 148–156
4. Nielsen, J.C., Bottcher, M., Nielsen, T.T., Pedersen, A.K. and Andersen, H.R. Regional
myocardial blood flow in patients with sick sinus syndrome randomized to long-
term single chamber atrial or dual chamber pacing—effect of pacing mode and
rate, J Am Coll Cardiol, 35, 2000, pp. 1453–1561
5. Skalidis, E.I., Kochiadakis, G.E., Koukouraki, S.I. et al. Myocardial perfusion in patients with
permanent ventricular pacing and normal coronary arteries, J Am Coll Cardiol,
37, 2001, pp. 124–129
6. Prinzen, F.W., Cheriex, E.C., Delhaas, T. et al. Asymmetric thickness of the left ventricular
wall resulting from asynchronous electric activation: a study in dogs with
ventricular pacing and in patients with left bundle branch block, Am Heart J,
130, 1995, pp. 1045-1053
7. Karpawich, P.P., Justice, C.D., Cavitt, D.L., Chang, C.H., Lee, M.A., Dae, M.W., Langberg, J.J. et al.
Effects of long-term right ventricular apical pacing on left ventricular perfusion,
innervation, function and histology, J Am Coll Cardiol, 24, 1994, pp. 225–232
8. Karpawich, P.P., Justice, C.D., Cavitt, D.L. and Chang, C.H. Developmental sequelae of fixed-
rate ventricular pacing in the immature canine heart:an electrophysiologic,
hemodynamic, and histopathologic evaluation, Am Heart J, 119, 1990, pp. 1077–
1083
9. Van Oosterhout, M.F., Prinzen, F.W., Arts, T. et al. Asynchronous electrical activation induces
asymmetrical hypertrophy of the left ventricular wall, Circulation, 98, 1998, pp.
588–595
10. Thambo, J.B., Bordachar, P., Garrrigue, S. et al. Detrimental ventricular remodeling in
patients with congenital complete heart block and chronic right ventricular
apical pacing, Circulation, 110, 2005, pp. 3766–3772
11. Adomian, G.E. and Beazell, J. Myofibrillar disarray produced in normal hearts by chronic
electrical pacing, Am Heart J, 112, 1986, pp. 79–83
12. Nahlawi, M., Waligora, M., Spies, S.M. et al. Left ventricular function during and after right
ventricular pacing, J Am Coll Cardiol, 4, 2004, pp. 1883–1888

502
 Quantitative Methods Inquires

13. Boucher, C.A., Pohost, G.M., Okada, R.D. et al. Effect of ventricular pacing on left ventricular
function assessed by radionuclide angiography, Am Heart J, 106, 1983, pp. 1105–
1111
14. Tse, H.-F., Yu, C., Wong, K.-K. et al. Functional abnormalities in patients with permanent
right ventricular pacing, J Am Coll Cardiol, 40, 2002, pp. 1451-1458

1
Acknowledgments:
This study was supported entirely by the Romanian Ministry of Education and Research through The National
University Research Council (CNCSIS) which we gratefully acknowledge.

2
Dr. Oana Stancu, MD-Internal Medicine and Cardiology, PhD, Lecturer
- Graduated at University of Medicine and Pharmacy”Carol Davila”, Bucharest, 1990;
- Assistant professor - Internal Medicine at University of Medicine and Pharmacy”Carol Davila”-1995
- PhD- 2002;
- Specialist - Cardiology 2004
- since 2005 - lecturer at University of Medicine and Pharmacy”Carol Davila”, Bucharest;
- since 1995 - MD at the Saint John Clinical and Emergency Hospital, Department of Internal Medicine and
Cardiology
Training in Pacing and Electrophysiology :
• 1996- training in cardiac pacemakers implants VVI,VDD,DDD-Military Hospital-Bucharest, in Cardiac
Surgery Department;
• 1998,1999-Training in pacing and electrophysiology, in Austria, Wien, Wilheminenspital;
• 2000-training in pacing and electrophysiology in Austria, Wien, Allgemeines Krankenhaus;
2001-Training in pacing and electrophysiology, Biotronik Company, Berlin, Germany.

3
Dr.Teodorescu Iulia- Cardiologist MD, PhD, researcher;
- graduated at University of Medicine and Pharmacy ”Carol Davila”, Bucharest, 1994;
- working by contest in Saint John Clinical and Emergency Hospital, Bucharest from 1995 in Internal Medicine
Department;
- researcher from 1995;
- PhD from 2002;
- cardiac echography classes, pacing and electrophysiology preoccupations;
• 1995-training in periphery angiography - National Institute of Cardiology Bucharest -Catheters and
Angiography Department
• 1996-training in cardiac pacemakers implants VVI, VDD, DDD - Military Hospital - Bucharest, in Cardiac
Surgery Department;
• 1998, 1999-Traing in pacing and electrophysiology offered by BIOTRONIK, in Austria, Wien,
Wilheminenspital; pacing and electrophysiology preoccupations;
• 2000-Training in pacing and electrophysiology offered by BIOTRONIK, in Austria, Wien, Allegemeines
Krankenhaus;
• 2001-training in pacing and electrophysiology offered by BIOTRONIK, Germany, Berlin

4
Emanuel Radu, MD, Researcher, Saint John Clinical and Emergency Hospital, Bucharest
- Graduated at University of Medicine and Pharmacy, Bucharest ”Carol Davila”;
- Working by contest in Saint John Clinical and Emergency Hospital, Bucharest, from 2000, Angiography
Deparment;
- Specialist in Cardiology since 2008

5
Zara Octavian Dumitru, MD, Researcher,resident in training in Interventional Cardiology,
- Graduated at „Vest University”-Vasile Goldis-1998;
- Working by contest in Saint John Clinical and Emergency Hospital, Bucharest, from 2000, Angiography
Deparment;
- resident in training in Cardiology-2002

6
Codification of references:
Saxon, L.A., Stevenson, W.G., Middlekauff, H.R. and Stevenson L.W. Increased risk of progressive
[1] haemodynamic deterioration in advanced heart failure patients requiring permanent
pacemakers, Am Heart J, 125, 1993, pp. 1306–1310
Alpert, M.A., Curtis, J.J., Sanfelippo, J.F. et al. Comparative survival after permanent ventricular and
dual chamber pacing for patients with chronic high degree atrioventricular block with
[2]
and without preexisting congestive heart failure, J Am Coll Cardiol, 7, 1986, pp. 925–
932

503
 Quantitative Methods Inquires

Rosenqvist, M., Isaaz, K., Botvinick, E.H., et al. Relative importance of activation sequence compared to
[3] atrioventricular synchrony in left ventricular function, Am J Cardiol, 67, 1991, pp. 148–
156
Nielsen, J.C., Bottcher, M., Nielsen, T.T., Pedersen, A.K. and Andersen, H.R. Regional myocardial blood
flow in patients with sick sinus syndrome randomized to long-term single chamber
[4]
atrial or dual chamber pacing—effect of pacing mode and rate, J Am Coll Cardiol, 35,
2000, pp. 1453–1561
Skalidis, E.I., Kochiadakis, G.E., Koukouraki, S.I. et al. Myocardial perfusion in patients with permanent
[5] ventricular pacing and normal coronary arteries, J Am Coll Cardiol, 37, 2001, pp. 124–
129
Prinzen, F.W., Cheriex, E.C., Delhaas, T. et al. Asymmetric thickness of the left ventricular wall
resulting from asynchronous electric activation: a study in dogs with ventricular
[6]
pacing and in patients with left bundle branch block, Am Heart J, 130, 1995, pp. 1045-
1053
Karpawich, P.P., Justice, C.D., Cavitt, D.L., Chang, C.H., Lee, M.A., Dae, M.W., Langberg, J.J. et al. Effects of
[7] long-term right ventricular apical pacing on left ventricular perfusion, innervation,
function and histology, J Am Coll Cardiol, 24, 1994, pp. 225–232
Karpawich, P.P., Justice, C.D., Cavitt, D.L. and Chang, C.H. Developmental sequelae of fixed-rate
[8] ventricular pacing in the immature canine heart:an electrophysiologic, hemodynamic,
and histopathologic evaluation, Am Heart J, 119, 1990, pp. 1077–1083
Van Oosterhout, M.F., Prinzen, F.W., Arts, T. et al. Asynchronous electrical activation induces
[9] asymmetrical hypertrophy of the left ventricular wall, Circulation, 98, 1998, pp. 588–
595
Thambo, J.B., Bordachar, P., Garrrigue, S. et al. Detrimental ventricular remodeling in patients with
[10] congenital complete heart block and chronic right ventricular apical pacing,
Circulation, 110, 2005, pp. 3766–3772
Adomian, G.E. and Beazell, J. Myofibrillar disarray produced in normal hearts by chronic electrical
[11]
pacing, Am Heart J, 112, 1986, pp. 79–83
Nahlawi, M., Waligora, M., Spies, S.M. et al. Left ventricular function during and after right ventricular
[12]
pacing, J Am Coll Cardiol, 4, 2004, pp. 1883–1888
Boucher, C.A., Pohost, G.M., Okada, R.D. et al. Effect of ventricular pacing on left ventricular function
[13]
assessed by radionuclide angiography, Am Heart J, 106, 1983, pp. 1105–1111
Tse, H.-F., Yu, C., Wong, K.-K. et al. Functional abnormalities in patients with permanent right
[14]
ventricular pacing, J Am Coll Cardiol, 40, 2002, pp. 1451-1458

504
 Quantitative Methods Inquires

DIFFERENT APPROACHES USING THE NORMAL AND THE

EXPONENTIAL DISTRIBUTION IN THE EVALUATION
OF THE CUSTOMER SATISFACTION

Antonio LUCADAMO1
PhD, Assistant Researcher, Tedass,
University of Sannio, Benevento, Italy

E-mail: [email protected]

Giovanni PORTOSO2
PhD, Associated Professor, SEMEQ Department, Faculty of Economics,
University of Eastern Piedmont “A. Avogadro”, Novara, Italy

E-mail: [email protected]

Abstract: The Customer Satisfaction is generally evaluated using the data collected with
questionnaires. The data are organized on an ordinal scale and, for this reason, it’s convenient
to transform them in pseudo-interval support. The psychometric methods used for this
transformation generally hypothesize that the latent variable has a normal distribution.
Sometimes, particularly when the frequencies are concentrated on the left extreme or on the
right extreme of the distribution, this assumption brings to preposterous results. In these cases
the use of other types of distribution, as, for example the exponential distribution, is preferable.
In this paper we show how the results of a survey can change using the normal distribution,
the exponential distribution or the two distributions alternatively. We use, in fact, the results
coming from the different transformations, to apply a multilevel model.

Key words: customer satisfaction; normal distribution; exponential distribution; multilevel

models

1. Introduction

One of the problem of the Customer Satisfaction is the quantification that converts
on a metric scale the judgements about services or products. A simple technique is the so-
called “direct quantification”: this technique hypothesizes that the modalities of a qualitative
character are at the same distance, but this hypothesis is not respected in many situations
(Marbach, 1974). For this reason it is preferable to use an alternative technique, the “indirect
quantification”, that consists in assigning real numbers to the categories of the qualitative
variable. In this type of quantification the numbers are not equidistant but they depend on a
latent variable. Different measurement techniques have been developed during the years
(Thurstone, 1925, Guilford, 1936, Torgenson, 1958) based on the hypothesis that the model

505
 Quantitative Methods Inquires

is normally distributed. This assumption can be realistic in a psychometric field, but it is not
always valid in the Customer Satisfaction, especially if the judgements are all extremely
positive or extremely negative. More recent techniques have been proposed, based for
example on the use of logit and probit models, on structural equation models and so on. In
next section we introduce the psychometric quantification, underlining the problems that can
arise in some situations; then we show how the use of another kind of quantification can
solve these pitfalls and, in the following paragraphs we propose the use of a combined
technique, showing the results that we obtain on real data.

2. The psychometric quantification

In the psychometric quantification, the modalities xi ( i = 1,2,… , r ) of a qualitative

variable X , are associated to the values of a quantitative latent variable Z , normally
distributed. Let F (i ) be the cumulative relative frequency, corresponding to xi and let
Φ −1 [ F (i)] the inverse of the cumulative distribution function, the quantile zi associated to
xi can be expressed as zi = Φ −1[ F (i)] . To obtain the new scores, we simply calculate the
expected values E ( Z i ) over all the X variables in the data-set. The assumption of the
normal distribution, when the frequencies are prevalently on the left extreme or on the right
extreme of the distribution, leads to strange results. In fact the scores will be negative if the
modalities are almost on positive side and vice-versa (the results in Table 1 can help to
understand the situation) (Portoso, 2003a).

Table 1. Quantification with the normal distribution

of the judgements given on two different services
Judgements Frequencies of the first service Frequencies of the second service
Very negative 350 10
Negative 80 20
Indifferent 40 40
Positive 20 80
Very positive 10 350
Totals 500 500
Expected quantile 0.0729 -0.0729

It is easy to see that the first service has many negative judgements, so the
frequencies are prevalently on the left side of the distribution, but the expected quantile has
a positive value. For the second service there is instead the inverse situation, in fact the
frequencies are on the right side, but expected value of the quantile is negative.
This incongruity leads to use a distribution that could better express, in a numerical
way, the categorical variables characterized by this particular structure. The exponential
distribution seems to be the right solution.

3. The exponential quantification

In this section we show how to determine a quantification based on the negative

and on the positive exponential distribution. Before introducing the new procedure, it’s
necessary to describe briefly the two cited distributions.

506
 Quantitative Methods Inquires

3.1. The negative exponential distribution

Let consider
⎧ψ ( z ) = exp(− z ) if (0 ≤ z ≤ ∞)
⎨ (1)
⎩ ψ ( z) = 0 otherwise
where Z is a quantitative variables.

It can be assumed as the relative density function, in fact:

∞ ∞

∫ψ ( z )dz = ∫ exp(− z )dz = 1

0 0
(2)

The mean and the variance are defined as follow:

∞ ∞ ∞
E ( Z ) = ∫ zψ ( z )dz = ∫ z exp(− z )dz = [− z exp(− z )] − ∫ − exp(− z )dz = 1
∞
0 (3)
0 0 0

∞
Var ( Z ) = ∫ z 2 exp(− z )dz − [ E ( Z )] 2 = 2 − 12 = 1 (4)
0

The variable can be then standardized in the following way:

S = ( Z − 1) = Z − 1 (5)
with
⎧ f ( s) = exp(− s − 1) if (−1 ≤ s ≤ ∞)
⎨ (6)
⎩ f (s) = 0 otherwise

This is a relative frequency density function, in fact:

+∞

∫ exp(− s − 1) = 1
−1
(7)

The cumulative distribution function is:

⎧ s

⎪Ψ ( s ) = ∫ exp(−t − 1)dt = 1 − exp(− s − 1) if ( −1 ≤ s ≤ ∞ )

⎨ −1
(8)
⎪ Ψ ( s ) = 0 otherwise
⎩

3.2. The positive exponential distribution

Let consider
⎧ψ ( y ) = exp( y ) if (−∞ ≤ y ≤ 0)
⎨ (9)
⎩ ψ ( y) = 0 otherwise

507
 Quantitative Methods Inquires

that can be assumed as the relative density function, in fact:

0 0

∫ψ ( y)dy = ∫ exp( y)dy = 1

−∞ −∞
(10)

The mean and the variance are defined as follow:

0 0 0
E (Y ) = ∫ yψ ( y)dy = ∫ y exp( y)dy = [ y exp( y)]−∞ − ∫ exp( y)dy = −1
0
(11)
−∞ −∞ −∞

0
Var (Y ) = ∫y exp( y )dy − [ E (Y )]2 = 2 − (−1) 2 = 1
2
(12)
−∞

The variable can be then standardized in the following way:

P = (Y − (−1)) = Y + 1 (13)

The cumulative distribution function of P is :

⎧ p
⎪Ψ ( p ) = ∫ exp(t − 1)dt = exp( p − 1) if (−∞ ≤ p ≤ 1)
⎨ −∞
(14)
⎪ Ψ ( p) = 1 otherwise
⎩

3.3. The quantification

To build the scores, both for the negative exponential distribution and for the
positive one, it’s necessary to consider the relative frequencies f (i ) and the cumulative
relative ones F (i ) .
In this way we can define the following quantity (empirical distribution of
cumulative frequencies):

G (i ) = F (i − 1) + f (i ) / 2 i = 1, 2,… , r (15)

If we consider the negative exponential distribution, we can compare formula (6)

and (15) and we obtain the standardized quantile:
s i = −1 − ln[1 − G (i )] (16)

The same procedure can be applied for the positive distribution, using formula (13)
and (15); in this case we will obtain the standardized quantile in the following way:

pi = 1 + ln[G (i )] (17)

508
 Quantitative Methods Inquires

To verify the importance, in particular situations, of using the exponential

distribution, instead than the normal distribution, we can consider the value in table 2 in
which there are the absolute frequencies about the judgments given to 8 different services
by 500 judges.

Table 2. Absolute frequencies of the judgements given to 8 different services

Services
Judgments A B C D E F G H
Very negative 496 470 20 180 20 10 2 1
Negative 1 16 50 50 120 20 4 0
Indifferent 1 8 360 40 220 40 6 0
Positive 1 4 60 50 120 350 96 0
Very positive 1 2 10 180 20 80 392 499
Total 500 500 500 500 500 500 500 500

In this table we can note that the services A and B received many negative
judgments, the services F, G and H many positive judgments and the services C, D and E
had a quasi-symmetric distribution.
This table is important to understand what happens when we apply the different
kinds of quantification.
The results are shown in table 3.

Table 3. Quantiles associated to the relative cumulate frequencies centred on every

judgment category in the hypothesis of exponential and normal distribution
JUDGMENTS
Expected
Services Very Negative Indifferent Positive Very quantile
negative positive
Exp neg. -0.315 3.962 4.298 4.809 5.908 -0.274
A
Norm -0.010 2.457 2.576 2.748 3.090 0.012
Exp neg. -0.385 2.124 2.912 3.828 5.215 -0.177
B
Norm -0.075 1.705 2.054 2.409 2.878 0.047
Exp neg. -2.912 -1.408 0.307 0.917 0.990 0.093
C
Norm -2.054 -1.341 0 1.405 2.326 -0.001
Exp neg. -0.802 -0.472 -0.307 -0.108 0.714 -0.114
D
Norm -0.915 -0.228 0 0.228 0.915 0
Exp neg. -0.980 -0.826 -0.307 0.833 2.912 -0.056
E
Norm -2.054 -0.994 0 0.994 2.054 0
Exp pos. -3.605 -2.219 -1.303 0.287 0.917 0.082
F
Norm -2.326 -1.751 -1.282 -0.025 1.405 -0.012
Exp pos. -5.215 -3.828 -3.017 -1.120 0.502 0.091
G
Norm -2.878 -2.409 -2.097 -1.175 0.274 -0.067
Exp pos. -5.908 -5.215 -5.215 -5.215 0.309 0.296
H
Norm -3.090 -2.878 -2.878 -2.878 0.003 -0.004
General mean 0.002

Here we can see that, in a Customer Satisfaction analysis, when the frequencies
are very high for judgments extremely positive or extremely negative, the use of the normal
distribution is not an appropriate way to effectuate the quantification. Using the exponential
distribution leads to better results, in fact we can see that for the services A and B that
presented value extremely negative, we have that the expected value of the quantile is

509
 Quantitative Methods Inquires

negative if we use the negative exponential distribution, while using the normal
quantification it will be positive. For the services C, D and E there are no substantial
differences between the use of the normal or of the exponential distribution, but the first one
seems to be preferable; for this services in fact we had a symmetric distribution. For services
G and H instead, the calculation of the expected quantile shows that the use of positive
exponential distribution leads to positive values, while using the normal distribution we will
have negative values.
Of course the observation of the expected quantile can not be the only instrument
to decide if considering the normal distribution or the exponential distribution as latent
variable, but we need an indicator that could help in the choice. A possible solution is given
in Portoso (2003b) that introduces an useful index to decide which kind of distribution is
better to apply in the different situations.

3.4. The EN index

The EN index is an indicator that assumes values between -1 and +1. The value -1
is assumed when all the frequencies are associated to the first modality (in this case we have
maximum negative concentration), while when there is maximum positive concentration the
value assumed by the index will be +1. If the frequencies are balanced in a symmetric way
then the EN index will be equal to 0. The index has the following formulation:
r/2
EN = ∑ ( f r −i +1 − f i )(r − 2i + 1) /(r − 1) (18)
i =1

where r is the number of modalities and if they are odd the value r/2 is round off to the
smaller integer while f i are, as already stated, the relative frequencies associated to the
modality i, f r −i +1 are the frequencies associated to the opposite modality and r − 2i + 1 is
the difference between the position of the two opposite modalities.
An alternative formulation of the index can be the following:
r −1
EN = 1 − 2∑ F (i ) /(r − 1) (19)
i =1

that presents some similarities with the Gini index and where F (i ) have already been
defined as cumulative relative frequency and r is the number of modalities of the qualitative
variables. If the value of the index EN is close to 0, the use of normal distribution doesn’t
generate any problems, but if the absolute value of this index grows then the use of
exponential distribution can lead to better results. The problem is to define a threshold to
decide which distribution is better to apply. Portoso, with empirical attempts, showed that a
value of the EN bigger than 0.2 in absolute value, indicates that the use of the exponential
distribution is preferable to the normal one. In the following sections we first introduce
briefly the multilevel models and then we verify what happens to the results of an analysis
using the different kinds of quantification.

510
 Quantitative Methods Inquires

4. The multilevel models

Multilevel models suppose that in a hierarchical structure, the upper levels can
influence the lower ones (Snijders, Bosker 1999). The basic model is the so called empty
model defined as follows:

Yij = γ 00 + U 0 j + Rij (20)

In this formula there is a dependent variable Yij given by the sum of a general

mean ( γ 00 ), a random group effect ( U 0 j ) and a random individual effect ( Rij ). In this way
the variability is divided in two parts: in fact, in this model it’s assumed that the random
variables U 0 j and Rij are mutually independent, normally distributed with zero mean and

variances equal to τ and σ . The total variance is then the sum of the two variances and
2 2

we can compute the intra-class correlation coefficient:

ρ = τ 2 /(τ 2 + σ 2 ) (21)

If this coefficient is significant, it is possible to effectuate a Multilevel Analysis (Hox,

2002). A first model is the Random Intercept Model that can be defined as follows:

Yij = β 0 j + β1 xij + Rij with β 0 j = γ 00 + U 0 j (22)

In the equation (20) if we consider the j subscript for the coefficient β1 we will
have the Random Slopes Model. In this case too we can see that there is a fix effect ( γ 10 )
and a random ones ( U 1 j ).

Yij = β 0 j + β1 j xij + Rij with β 0 j = γ 00 + U 0 j and β1 j = γ 10 + U 1 j (23)

5. A case study

The application concerns a survey about Patient Satisfaction. The patients answered
to 30 items relative to the services received during the staying in the hospital. They gave a
score between 1 and 7 and furthermore they furnished information about the gender, the
age and the education. To apply a multilevel model we need a variable relative to the
second level and this is the experience of the head physician of the different wards. The aim
is to verify if the Customer Satisfaction (CS) depends on the different variables and, above
all, if the different quantifications leads to dissimilar results. For this reason we adopt the
Normal Quantification, the Exponential Quantification and a Mixed Quantification (Normal
or Exponential). The first two quantifications have already been illustrated, instead the third
one is based on the use of the EN index. We use in fact the normal quantification for the
items that have the EN index lower than a fixed threshold and the exponential distribution
for the items with EN larger than the threshold. Furthermore, to compute the new scores, we
used a geometric mean for the exponential quantifications because of lower sensitivity to

511
 Quantitative Methods Inquires

extreme values and an arithmetic mean for the normal distribution. In Table 4 we show the
number of the items transformed using the two distributions, according to the different
thresholds, arbitrarily assumed and considering that it is possible a larger series of values.

Table 4. Number of the two distributions used according to the different thresholds
Threshold 0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1
N° of exponential 30 29 26 25 22 17 9 1 0
N° of normal 0 1 4 5 8 13 21 29 30

For all the criterions we then compute the overall CS as the sum of the new value
that every individual has for the 30 items. In the building of the model, the only significant
variable for the individual level is the age and the model that we adopt is a Random
Intercept Model, so we can write:

CS ij = β 0 j + β1 Ageij + Rij with β 0 j = γ 00 + γ 01 ( Exp j ) + U 0 j (24)

The results that we obtain are reported in Figure 1.

Experience Age

0.75 0.18
0.70 0.17
0.16
0.65 0.15
0.60 0.14
0.55 0.13
0.12
0.50 0.11
0.45 0.10
0.09
0.40
0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1
0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1
Threshold
Threshold

Figure 1. Coefficients of the two explicative variables with the different thresholds

We can note that the coefficients relative to the experience of the head physician
( γ 01 ) and to the age of the patients ( β1 ) are both positive, so the CS is higher for older
patients and for people that were nursed in department with expert doctors. Furthermore
they increase if we consider an EN threshold that goes from 0 to 0.2 and then they both
decrease considerably when the threshold is higher than 0.2, reaching a minimum by using
only the normal as latent variable. Moreover they are all statistically significant and there are
no substantial differences in the values of the t-ratio. The value of the EN = 0.2 is the value
that Portoso (2003b) indicated as critical for the choice between exponential and normal
distribution.

6. Considerations and perspectives

In the Customer Satisfaction or in the evaluation of other services there is a

quantification problem that can not be solved using the direct quantification, because it
doesn’t answer to the reality. The use of the indirect quantification, with the assumption of a
continuous latent distribution, is in this case preferable, but the choice can not be always the
normal distribution. Using its standardization, the exponential distribution, negative or

512
 Quantitative Methods Inquires

positive, has been assumed as an alternative to the normal when this one is not appropriate.
The exponential distribution assures results that are more consistent with the shape of the
empirical distribution and, furthermore, it guarantees distances between the modalities more
adherent to the psychological continuum with which the judgments are expressed. The
problem about the choice of the right distribution was discussed in an empirical way in a
previous work; in this paper, the results that we obtain, introducing also a second step of the
analysis, confirm the idea of using the exponential distribution instead than the normal one
when the EN index is higher than 0.2 or smaller than -0.2. Obviously these are only results
that comes from a restricted number of analyses and the definition of the threshold for the
choice between normal and exponential distribution must be studied deeply. Furthermore
some other indexes could be proposed and not only the use of normal and exponential
distribution must be taken into account; our proposal is in fact to consider, in next works, the
use of other distributions too.

References

1. Guilford, J.P. Psychometrics Methods, McGraw-Hill, New York, 1936, pp. 255-258
2. Hox, J.J. Multilevel Analysis. Techniques and Applications, Lawrence Erlbaum
Associates, 2002
3. Marbach, G. Sulla presunta equidistanza degli intervalli nelle scale di
valutazione, Metron, Vol. XXXII: 1-4, 1974
4. Portoso, G. La quantificazione determinata indiretta nella customer satisfaction:
un approccio basato sull’uso alternativo della normale e
dell’esponenziale, Quaderni di dipartimento SEMeQ, 53, 2003a
5. Portoso, G. Un indicatore di addensamento codale di frequenze per variabili
categoriche ordinali basate su giudizi, Quaderni di dipartimento SEMeQ,
66, 2003b
6. Snijders, T. A. B. and Bosker R. J. Multilevel Analysis. An introduction to basic and
advanced multilevel modelling, SAGE Publications, 1999
7. Thurstone, L. L. A method of scaling psycological and educational tests, J. Educ.
Psychol., 16, 1925, pp. 443-451
8. Torgenson, W.S. Theory and Methods of Scaling, Wiley, New York, 1958

1
Antonio Lucadamo, Philsophy Doctor in Statistics with the thesis “Multidimensional Analysis for the definition of
the choice set in Discrete Choice Models”- department of Mathematics and Statistics, University of Neaples
“Federico II”. From 2007-2009 he had a post-doc research fellowship in Statistics about the Customer Satisfaction –
department SEMEQ, University of Eastern Piedmont, Novara. Currently he has a post-doc research fellowship about
the use of multivariate statistical methods in the spectrometry –Tedass, University of Sannio, Benevento. His
research field are the Discrete Choice Models (Logit and Probit), the Multidimensional Analysis of Data (Principal
Component Analysis, Cluster Analysis), the Multivariate Analysis for the definition of the sample size in clinical trials,
the Rasch Analysis, the Quantification Methods and the Multilevel Regression for the evaluation of Customer
Satisfaction.

2
Graduated in Economics from the University of Bari in 1965 by a vote of 110 cum laude. Assistant Professor since
1971. In 1982 he became Associate Professor at the Faculty of Economics of Bari. He then taught at the Faculty of
Economics in Turin. In the nineties, with 4 other colleagues, he formed the Faculty of Economics of Novara, where
he currently teaches. Declared fit to full professor, he is not called by the Board of Faculty of Economics of Novara
for lack of quorum. He has lectured in Anthropometry, Statistics, Statistical Methodology, Statistical Quality Control,
Multivariate Statistical Analysis, Statistical Sampling, Theory of Probability and Mathematical Statistics, Market
Research. His publications cover the methodological aspects and applications developed in the field of
demographics and the financial sector, anthropometry, statistical analysis of the results of gambling, use of latent
variables in the customer satisfaction, normalization of indices of association.

513
 Quantitative Methods Inquires

THE IMPACT OF FINANCIAL CRISIS

ON THE QUALITY OF LIFE

Lindita ROVA
MSc, Lecturer, Economic Department,
“Eqrem Cabej” University, Gjirokastra, Albania

E-mail: [email protected]

Romeo MANO
PhD Candidate, Department of Mathematics, Physics and Informatics,
“Eqrem Cabej” University, Gjirokastra, Albania

E-mail: [email protected]

Abstract: The quality of life is a relatively new concept, which is continually changing and for
which there is not yet a wholly satisfactory definition. The quality of life involves human, social-
economic and health characteristics. The manifold nature of the quality of life led to the
development of various patterns for measuring it. The quality of life is determined by subjective
and objective indices and this allows us to have a clearer overall picture of it.
The aim of this research paper is to measure the standard of living based on six main levels
determined by the European Foundation for the Improvement of Living and Working
Conditions-Eurofound: 1. Employment; 2. Total income; 3. Family & Home; 4. Social life &
Community involvement; 5. Medical care & health security insurance and 6. Knowledge &
education
A very interesting part of the paper is the research into the impact of economic crisis on the
quality of life. The study is based on a questionnaire which is filled by 200 persons living in
Gjirokastra and Saranda. The analysis of the results from the questionnaire is carried out
employing the logistic regress method.

Key words: quality of life; economic crisis; welfare; objective indices; subjective indices;
logistic regression method; Albania

1. What is the concept of quality of life; a historical overview

“People are the real wealth of a nation. Thus, the main real aim
of development is to provide the basis for an environment where
people could be able to lead a long, healthy and creative live.
Human development brings about an increase in the number of
possible alternatives offered to the individual, of which the most
important are to live longer and healthier and to get used to
enjoying a respectably high standard of living. Other alternatives
include political freedom, human rights and self-respect.”
UNDP 1990
The quality of life is a relatively modern concept, which has undergone significant
changes, and for which, there does not exist a universally excepted definition. The quality of

514
 Quantitative Methods Inquires

live involves human, social, economic and health characteristics (Lamau, 1992; Treasury
Board of Canada Secretariat, 2000).
Historically speaking the quality of life as a concept appeared first in the 50s
referring to a high living standard in the new consumer society just created. The way in
which it was measured was the accumulation of wealth (money, new cars, houses, furniture,
modern electrical appliances etc). Later, in the 60s, the concept extended to include
elements such as education, health care, economic and technological development as well
as the protection of the Free World (Fallowfield, 1990). During the 70s, times in which, the
development schemes failed to improve the living standards of the poor strata of the society,
there was a heated long-running debate over the real goals of the economic development
as well as the political goals supposed to be achieved by the developed countries. Thus, the
concept of basic human needs as the fundamental estimated parameter of development
schemes was brought to light (Nagamine, 1981:1).
This perception, highlighting the human basic needs and reduction of poverty did
not except economic growth as the only goal of economic development. On the contrary a
special interest was shown in the quality of life of the poor strata of the society and the
distribution of income. Meeting the basic needs has often been identified with those needs
providing the bases for a living standard in accordance with human dignity. Put otherwise,
this means that different generally accepted patterns should be studied considering food,
clothing, housing, health, education, social safety, working conditions and human freedom
(Moon, 1991:5).
In the 60s and 70s the concept of wellbeing (as a subjective perception) was
extended to include happiness or life enjoyment as a new dimension in asserting the quality
of life. For many researchers it was considered to be the most important (Argyle, 1996) and
later, fronted by Amartia Sen, several new elements were added to it, to determine human
development. In Sen’s research (1993) individual ability was introduced as a concept making
for the first time the connection between the quality of live and people’s ability to be
involved in important activity mutually beneficial of both individual and society. In the 90s,
by putting these theories together one could talk about human development, which in fact is
a multi dimensional concept.
The fundamental idea of human development is that wellbeing is a crucial
parameter of development and that the individual is the bases of every level of this
development.
The quality of life as a concept involves several subjective and objective elements
and every study on the quality of life should consider all these elements. The subjective
space refers to the wellbeing and pleasures one gets from the environment in which one
lives, whereas the objective space refers to the pleasures of the individual related to social
political requirements; material wellbeing, social status, good medical conditions, etc (Noll,
1998).
Studies on the quality of life based only on objective indices cannot be complete.
For this reason, a serious study should analyze objective and subjective indices both from the
qualitative and quantitative point of view. According to European Foundation for the
Improvement of Living and Working Conditions (Eurofound) the quality of life is a clear
indication related to economic wellbeing in social context and has 6 main areas of study:
1. Employment
2. Total income
3. Family & Home
4. Social life & Community involvement
5. Medical care & health security insurance
6. Knowledge & education

515
 Quantitative Methods Inquires

2. Objective indices of the quality of life assessment.

This study emphasizes some of the objective parameters measuring wellbeing as a

component of the quality of life and will stop to analyze mainly its subjective elements.
Among the objective elements we should highlight the parameters universally accepted as
important in measuring economic growth for example:
1. Gross Domestic Product
2. Gross Domestic Product per capita
3. Investments as percentage of Gross Domestic Product
4. Average salary
5. Average income per capita after tax
6. Minimal pay (salary)

Employment & unemployment indicators:

1. Population (Active – Non active ) according to group age & Gender
2. Employment
3. Unemployment
4. Full-time, Part- time employment
5. Employment according to economic fields or occupation etc.

Parameters for measuring material goods & services:

1. Number of cars
2. Number of schools- primary & secondary schools
3. Number of hospitals, clinics etc.
4. Illiteracy rate
5. Number of pupils & students
6. Number of doctors
7. Number of computers etc.

Indicators for public expenses:

1. Public expenses as percentage of GDP
2. Public expenses for health care
3. Private expenses for health care
4. Expenses for old age pensions as percentage of GDP
5. Expenses for employment programs as percentage of GDP
6. Expenses for social welfare as percentage of GDP etc.

The combination of all these indicators can provide a clear picture over the level of
wellbeing and development of a country; it can help even to compare the countries among
themselves. But the quality of life comprises elements other than the above mentioned ones.
The most delicate and, therefore, the most important factor in the process of its evaluation is
undoubtedly the implication of subjective elements in it.

3. The subjective indices of quality of life assessment

The main goal of this study is precisely the assessment of the subjective elements
and their impact on determining the quality of life. A questionnaire was compiled for this
reason comprising several categories which takes place in this determination. Specifically in
the questionnaire is asked the evaluation of the interviewee for:
• Material wellbeing (income, purchasing power, housing)
• Wealth

516
 Quantitative Methods Inquires

• Political stabilization and security

• Family life
• Employment and work safety
• The connection between employment and personal life
• Culture and sports
• Social quality evaluation

The survey is conducted in the Southern Region of Albania including three districts:
Saranda, Gjirokastra and Tepelene. Initially we thought to carry out the survey based on 250
questionnaires. The questionnaire contains 38 questions and 221 persons were questioned
out of 250. The survey area was divided in several groups in order to have a complete cover
of the area according to the formula: 30% in rural areas, 30% at random choice (random
sample), 10% in private business, 10% in family environments, 10% employed in the private
sector, 10% employed in the public sector. The questionnaire was distributed in the whole
region according in the following way: 60% in Gjirokastra, 25% in Saranda and 15% in
Tepelena according to the number of inhabitants in the three districts respectively. The
questionnaire contains six questions related to the resent economic and financial crisis, in
order to study its link with the quality of life. Data processing was carried out based on
logistic regress method and here are some of the results.

a) Assessment of quality of life elements

The first part of the survey deals with material welfare as one the main elements of
determining the quality of life. Material welfare can be determined in two ways or according
to two principal view points, from the objective one (income level and their growth rate,
individual purchasing power etc) as well as the subjective one (the satisfaction he gains from
possessing material goods). Put otherwise, two individuals with the same purchasing power
ability judge their situation differently according to their own requirements. The outcome of
the survey is interesting: 47% of the interviewees feel relatively rich, 14% feel rich, 28%
relatively poor, 5% very rich and 6% very poor (figure 1).

Figure 1. The poverty level based on Figure 2. The basic needs fulfillment
incomes

517
 Quantitative Methods Inquires

The rate of meeting their basic needs divides the survey area into two parts: 51% of
them meet their basic needs adequately, 20% little and the same percentage much, 4% very
little and 5% meet their basic need too much. We think that the answers to the above
mentioned questions have a considerable margin of insincerity. That’s because the
Albanians are too proud to accept their level of poverty (graph nr 2).
Concerning employment the survey reveals that 60% of the interviewees were
employed, 7% unemployed, 5% retired (graph nr 3)
Concerning medical care in the region 30% of the interviewees asses it as
adequate, 18% as good, 17% as not good at all, 7% as very good (graph nr 4). Social
stability and safety were assessed to be as follows: 35% of relatively high level, 33%
relatively low level, 14% very low and only 3% judged it to be very high (graph nr 5)

Figure 3. Employment Figure 4. Health services quality

In graph number 6 represents the results of five questions concerning cultural life
and sports (question nr 13), the quality of social services such as education etc offered in the
region (question number 14), assessment on the environment related to greenery and
pollution (question 15), every day facilities such as roads, traffic, public transport system
(question 16) and as a syntheses of all the above mentioned questioned there is question
number 17: how would you assess the quality of your life.

Figure 5. Social stability and safety Figure 6. The results for questions 13,
14, 15, 16 and 17

518
 Quantitative Methods Inquires

Again the outcome of the survey is interesting: more than 50% think that the
cultural and sport life (Q13), the environment their live (Q 15) and the everyday facilities (Q
16) offer them little or too little satisfaction. The quality of the public services (Q 14) is
thought to be little or not at all satisfactory from 39% of the interviewees, averagely
satisfactory from 40% and 17% of them feel satisfied and very satisfied. As a synthesis of all
the above questions the interviewee is asked to express their feelings about their quality of
life (Q 17) and here we have contradictory results. 30% of them are minimally or not at all
satisfactory, 42% are rather satisfactory and the rest, which means 28% feels from
satisfactory to very satisfactory (graph nr 7).

Figure 7. Quality of life evaluation Figure 8. The world crisis information

b) The relation between the economic crisis and quality of life

The part of the survey which tends to show the assessment of the impact of the
economic crisis concerning the quality of life starts with the question: are you aware of the
resent world economic crisis? The survey reveals that 89% of the interviewees answered
positively, 7% of them answered negatively and 4% didn’t give an answer (graph nr 8)
Logistic regress applied to understand the link between question number 18 (Are
you aware of the current world economic crisis) as a reactive variable (the Y of the logistic
regress) and question number 21 (What has happened with the general price level) as a
predictive variable (the X of the logistic regress) shows an odd ratio of a very high level
(1.8099). This goes to show that according to the interviewees’ price increase from a base
level to another lever with one unit increases 1.8 times the possibility of crisis sensitivity on
the consumers part (table nr 1 logistic regress)
The link between question number 18 as a reactive variable and question 20 as a
predictive variable reveals an odd ratio lower than the unit (0.8867) which means that in
mean time there are no positive chances for an increase in the family income. On the
contrary it is expected to be in decline due to the impact of the economic crisis on the real
income.

519
 Quantitative Methods Inquires

Table 1. Logistic regression between Q18 and Q19-23

Logistic regression
Dependent Y Y
Method Enter
Sample size 207
Cases with Y=0 15 (7.25%)
Cases with Y=1 192 (92.75%)
Overall Model Fit
Null model -2 Log Likelihood 107.62585
Full model -2 Log Likelihood 99.65450
Chi-square 7.9714
DF 5
Significance level P = 0.1578
Coefficients and Standard Errors
Variable Coefficient Std.Error P
X19 0.0505 0.4062 0.9010
X20 -0.1203 0.2834 0.6712
X21 0.5933 0.4530 0.1904
X22 0.0630 0.4496 0.8886
X23 -0.9422 0.4775 0.0485
Constant 3.4529
Odds Ratios and 95% Confidence Intervals
Variable Odds Ratio 95% CI
X19 1.0518 0.4745 to 2.3317
X20 0.8867 0.5088 to 1.5451
X21 1.8099 0.7448 to 4.3982
X22 1.0650 0.4412 to 2.5710
X23 0.3898 0.1529 to 0.9938
Classification table
Actual group Predicted group Percent correct
0 1
Y=0 0 15 0.00 %
Y=1 0 192 100.00 %
Percent of cases correctly classified 92.75 %
The impact of the economic crisis on the interviewees’ income shows the following:
22.6% think that their income is the same in rapport to the crisis, 45.7% have a slight
decrease in their income, 27.1% have a considerable decrease and 4% have an increase in
their income. From the logistic regression analyses it appears that the difference in the
income level has no impact on crisis sensitivity because the odd ratio is near the unit.
The method of logistic regress was applied in our study in order to find the link
between question number 18 (Are you aware of the current world economic crisis) and the
questions 30 (Where do you live; in rural or urban area), question 32 (The type of housing;
owned or rented) and question 33(The educational level). All the three variables seem to
have had a considerable impact on the interviewees’ answers on the recent crisis. The most
considerable impact seems to be noticed on variable 32 (The type of housing) which means
that those families that live in rented accommodation, thus having a higher living cost, tend
to be more liable to recent crisis impact. Such a family feels the economic effect of the crisis
three times more than the families that live in their own accommodation (odd ratio 3.1373),

Table 2. Logistic regression between Q18 and Q30, 32, 33

Logistic regression
Dependent Y Y
Method Enter

520
 Quantitative Methods Inquires

Sample size 207

Cases with Y=0 16 (7.73%)
Cases with Y=1 191 (92.27%)
Overall Model Fit
Null model -2 Log Likelihood 112.65429
Full model -2 Log Likelihood 108.74713
Chi-square 3.9072
DF 3
Significance level P = 0.2717
Coefficients and Standard Errors
Variable Coefficient Std.Error P
X33 0.5772 0.3647 0.1134
X30 0.6319 0.6381 0.3220
X32 1.1434 1.0707 0.2856
Constant -1.4615
Odds Ratios and 95% Confidence Intervals
Variable Odds Ratio 95% CI
X33 1.7811 0.8715 to 3.6399
X30 1.8813 0.5386 to 6.5713
X32 3.1373 0.3847 to 25.5820
Classification table
Actual group Predicted group Percent correct
0 1
Y=0 0 16 0.00 %
Y=1 0 191 100.00 %
Percent of cases correctly classified 92.27 %

Of no little importance seems to be the difference in sensitivity to the crisis in the

rural areas compared to the urban ones. According to the study the possibility that families
living in the rural area feels the crisis is two times more than the families living in the urban
ones (odd ratio 1.8813). As well as this, the link between the education level and crisis
sensitivity due to a better understanding of things, shows that the increase in the education
level with one unit increases the crisis sensibility by 1.7811 times.
If we add the political impact on the above mentioned regress (question number
24: Are you personally interested in politics?) it is obvious that crisis sensitivity remains the
same whether one lives in the rural or urban area. The same can also be said for the
education level, but things are different when it comes to the kind of accommodation people
live in, case in which the crisis sensitivity appears to increase considerably (from 3.1373 to
4.0951). This means that the political impact of economic crisis on the families with rented
accommodation is more significant.

521
 Quantitative Methods Inquires

Table 3. Logistic regression between Q18 and Q30, 32, 33 and Q24
Logistic regression
Dependent Y Y
Method Enter
Sample size 207
Cases with Y=0 16 (7.73%)
Cases with Y=1 191 (92.27%)
Overall Model Fit
Null model -2 Log Likelihood 112.65429
Full model -2 Log Likelihood 95.69744
Chi-square 16.9568
DF 4
Significance level P = 0.0020
Coefficients and Standard Errors
Variable Coefficient Std.Error P
X33 0.4737 0.3882 0.2224
X30 0.5468 0.6664 0.4119
X32 1.4098 1.1208 0.2085
X24 -0.9250 0.2869 0.0013
Constant 1.7539
Odds Ratios and 95% Confidence Intervals
Variable Odds Ratio 95% CI
X33 1.6059 0.7503 to 3.4371
X30 1.7278 0.4680 to 6.3790
X32 4.0951 0.4552 to 36.8408
X24 0.3965 0.2260 to 0.6958
Classification table
Actual group Predicted group Percent correct
0 1
Y=0 0 16 0.00 %
Y=1 0 191 100.00 %
Percent of cases correctly classified 92.27 %

On the other hand the level of the interviewees’ interest in politics, which normally
expresses their level of trust in it and the alternatives offered by it, has a major impact on
their perception of the crisis. As the answers to question number 24 (Are you personally
interested in politics?) show a decrease in the interest in politics (odd ratio has a much lower
level than the unit, 0.3965) which means that the lower the interest in politics the lower the
sensitivity towards the crisis becomes.
As a conclusion we can say that politics and momentary political events such as
elections, political instability, and frequent rotation in office and frequent changes in
government increase the level of the sensitivity towards the economic crisis. The answer to
the question: how you forecast the crisis in the months to come 54.7% of the interviewees’
answered “I don’t know”, 20.4% of them think the situation is going to improve and 23.9%
think that it is going to deteriorate.

522
 Quantitative Methods Inquires

4. Conclusions

The above survey shows the following results:

1. In the quality of live assessment a very important part is played by subjective
indicators along side with the objective ones. Therefore it is necessary that they are
taken into consideration in its assessment. Of great importance in this context are
the methods applied to measure them. Certainly every method should be based on
surveys where the individuals express their subjective assessment concerning the
integral components of the quality of life.
2. These indices are relative and of national character related to culture, tradition and
the other peculiarities of each nation.
3. The quality of health care in the region is inadequate because 75% of the
interviewees value it from satisfactory to not good at all.
4. Cultural life and sports in the region is poor, this because 88% of the interviewees
give a negative answer to it, offering them little or no satisfaction at all.
5. Public services are beginning to improve. Consequently, there is a good attitude to
them. Therefore 55% of the interviewees assess themselves to be rather satisfied,
satisfied and very satisfied. Nevertheless, if it is thought that the requirements for the
public sector will be on the increase, the assessment on their quality will probably
change.
6. An increase in the amount of information and sensibility has been of great help to
the citizens to know and assess the economic crisis of the recent months. The
number of those interviewed with knowledge and perception on the crisis and its link
with income level, purchasing power level and the general price level is high. All
these factors increase the stress level and have a negative impact on the quality of
life.
7. Interest shown in politics, which indirectly measures the trust and commitment level
of broad masses in it, was low. 60% of the interviewees are little or not at all
interested in politics.
8. In this research paper it is not possible to state out if the quality of life has been on
the increase or otherwise because of the lack of the basis for comparison. This study
will provide the basis for a further research to compare the Albanian quality of live in
the future.

5. Recommendation

1. During the process of assessment of the quality of life, it is indispensable that decision
making authorities in cooperation with the civil society should take into account objective
indicators as well as subjective ones.
2. It is high time that human development in Albania was considered a priority as it is
nowadays in all developed countries. Besides the classical concept (economic growth) it
should be extended to include such new concepts as social, cultural, educational
development, increase security and qualitative public services.
3. In order to enable the achievement of satisfactory levels of human development the
quality of life should be increased. It is for this reason that an increased rate of
awareness is considered to be of paramount importance. The community should put

523
 Quantitative Methods Inquires

pressure on the decision making authorities so that they improve their policies aiming at
the betterment of all the afore-mentioned indicators.

References

1. Agresti, A. An introduction to Categorical Data Analysis, New York, Wiley, 2002

2. Argyle, M. The Social Psychology of Leisure, Penguin, London, 1996
3. Fallowfield, L. The Quality of Life: The Missing Measurement in Health Care, Souvenir Press
(E & A), London, 1990
4. Fox, J. Applied Regression Analysis, Linear Models and Related Methods,. Sage publication,
Thousand Oaks, CA,1997
5. Fox, J. Logistic Regression, Maximum-Likelihood Estimated and Generalized Models, York
Summer Programme in Data Analysis, May 2005
6. Lamau, M.L. The Idea of Quality of Life in the Health Field, The Quality of Life in the
Mediterranean Countries, First Mediterranean Meeting on Bioethics, Instituto Siciliano di
Bioetica, 1992, pp. 47-68
7. Mano, R. Regresi Logjistik dhe një zbatim në microfinance, Mikrotezë Master, Departamenti I
Matematikës, FSHN, UT, 2005
8. Moon, B. E. The Political Economy of Basic Human Needs, Cornell University Press, London,
1991
9. Nagamine, H. Human Needs and Regional Development, Nagoya, Japan, 1981
10. Noll, H.-H. Societal Indicators and Social Reporting: The international Experience, The
Quality-of-Life Research Center, Copenhagen, 1998
11. Sen, A. Capability and Well-being in the Quality of Life, Nussbaum, M.C. and Sen, A. (eds.).
Oxford University Press, 1993
12. * * * Easy-to-use statistical software, MedCalc, 2002
13. * * * Quality of Life – A Concept Paper. Defining, Measuring and Reporting Quality of Life
for Canadians, Treasury Boeard of Canada Secretariat, 2000

524
 Quantitative Methods Inquires

STATISTICAL INDICATORS FOR RELIGIOUS STUDIES:

INDICATORS OF LEVEL AND STRUCTURE1

Claudiu HERTELIU2
PhD, Assistant Professor, Department of Statistics and Econometrics,
University of Economics, Bucharest, Romania
(Co)Author of the books: Metode cantitative in studiul fenomenului religios (2009),
Strategia universităţii: Metodologie şi studii de caz (2007)

E-mail: [email protected], Web page: http://www.hertz.ase.ro

Alexandru ISAIC-MANIU
PhD, University Professor, Department of Statistics Econometrics
University of Economics, Bucharest, Romania
(Co)Author of the books: Abordarea Sase Sigma. Interpretari, controverse, proceduri
(2008), Proiectarea statistica a experimentelor (2006), Enciclopedia calitatii (2005),
Dictionar de statistica generala (2003)

E-mail: [email protected], Web page: http://www.amaniu.ase.ro

Abstract: Using statistic indicators as vectors of information relative to the operational status
of a phenomenon, including a religious one, is unanimously accepted. By introducing a system
of statistic indicators we can also analyze the interfacing areas of a phenomenon. In this
context, we have elaborated a system of statistic indicators specific to the religious field, which
highlights the manner of positioning, evolving, interfering and manifesting of this field in
society.

Key words: statistical indicators; religious studies; indicators of level; indicators of structure;
indicators’ descriptors; indicator card

Context

Statistic indicators systems have been build to be applied in various areas:

education, environment, etc. Depending on the destination of these systems, their complexity
and dimension can be smaller or larger. In the followings, we present a few descriptive
elements for a component of the system of statistic indicators proposed for the analysis of
the religious phenomenon. The diagram below presents the general context and structure of
the system of indicators for the analysis of the religious phenomenon. The indicators system
is structured in six sub-systems (some of them divided in other categories), and has multiple
approaches of the religious phenomenon: i) the use of different reporting coordinates, ii) the
analysis based on various aggregation/depth levels, iii) the use of various measurement
scales, iv) the definition of both primary (absolute as well as relative) and derived indicators,
etc.
Not all presented indicators are innovative. Some of them (e.g. the structure of the
population by declared religious affiliation) are generally used by most of the researchers,
and contribute to obtaining the general image of the pursued phenomenon.

525
 Quantitative Methods Inquires

Figure 1. The structure of indicators’ system for religious studies

This article is intended to detail the third sub-system of indicators: Indicator of level
and structure, the indicators of this sub-system are the following: Indicators of level an
structure. The indicators of this subsystem are the following:
1. Number of existing religious groups (NRG)
2. Structure of the population from viewpoint of the religious affiliation (RLG)
3. Minimum number of religious groups for the establishment of religious oligopoly
(OLIG)
4. The structure of the population from viewpoint of the active religious affiliation
(ACTRLG)
5. The balance of legally constituted families (LEGFAM)
6. The balance of multi-confessional families (MCONFAM)

Indicators’ descriptors

Indicator 1. Number of existing religious groups

Characteristics
Represents the total number of religious groups legally constituted and
Definition
recognized “as is” in a distinct geographical area (usually a country).
The indicator measures the level of flexibility and easiness in obtaining a
“license” for a certain communion. It reflects the degree of religious freedom,
Scope
compatibility and interest on which raises a certain communion within a
country.
Symbol NRG
Calculus method It counts the number of religious groups recognized by an authority.

526
 Quantitative Methods Inquires

NRG=max(i) (1.)
where:
Formula
i – the numbers of order in the list of confessions legally recognized by the
authorities
Necessary data The list of legal confessions as recognized by the authorities.
Data source The Ministry of Culture and Cults
From the data type point of view : absolute
From the measurement scale point of view: interval
Type
From the calculus point of view: primary
From the time evolution point of view: moment
Class/ category Indicators of level and structure
Aggregation level National, regional, continental, worldwide, etc
It characterizes the degree of attractiveness and freedom within a country or
Interpretation
region from the religious activities point of view.
One must consider the differentiation between the various phases a
Quality standards cult/church/movement must cover until the complete recognition or until the
stage of legally accepted or tolerated group.

Indicator 2. Structure of the population from viewpoint of the religious affiliation

Characteristics
The weights of religious institutions from the perspective of the number of self-
Definition
declared believers, in a certain geographical region.
The indicator is a classical one and maybe the most commonly used indicator
for the study of religious phenomena. Based on its results, international
Scope
comparisons can be made or it can characterize the existent type of “religious
market”.
Symbol RLG
It is calculated by relating the total number of persons affiliated to a religious
Calculus method denomination to the total population investigated. Facultative, the result can
be expressed as a percentage by multiplication with 100.
Bi
RLGi = ⋅ [100] (2.)
∑ i
B
Formula Where:
Bi – the number of believers affiliated to the o confession
∑B i - the volume of the investigated population
The number of believers for each existing confession in the studies population
Necessary data
(the C indicator within the indicators characterizing human resources).
Data source NIS3, Religious Institutions, Censuses4
From the data type point of view : relative
From the measurement scale point of view: continuous
Type
From the calculus point of view: derived
From the time evolution point of view: moment
Class/ category Indicators of level and structure
Local, by county, regional, national, continental, worldwide, by gender, by
Aggregation level
residence environments, etc.
We can also measure and make comparisons between the existing situation
Interpretation regarding religious affiliation in various geographical areas or in population
structures by various aggregation criteria (gender, education, nationality, etc).
In order to give a good interpretation of the results, one must also consider
the fact that at a certain moment in time, a person can belong to a single
church. Also, the denominator and the numerator of the fraction must be
Quality standards
specific to a population synchronized spatially and timely. Another issue is
given by the fact that this indicator is almost always based on the free
statement of the individuals.

527
 Quantitative Methods Inquires

Indicator 3. Minimum number of religious groups for the establishment of

Characteristics religious oligopoly
The number of confessions, calculated on a descending sorted scale
depending on the ratios obtained, assuring a limit percentage ( ε ) of the total
Definition
number of believers. The limit percentage can be set, for example, to a level
of 80%.
The indicator measures the degree of saturation/ diversification existent on
Scope
the “religious market” in a certain location.
Symbol OLIG
The weights obtained for each confession are determined and are cumulated
Calculus method increasingly. OLIG is the rank of the confession for which the value of ε is
outran for the first time.
OLIG = min(i ) Fi ∈ B; B = { Fi ≥ ε } (3.)

Formula Where:
Fi – are the cumulative frequencies sorted decreasingly
ε - is the minimum level set for constituting oligopoly.
Necessary data The number of believers for each confession in the studied population.
Data source NIS, Religious Institutions
From the data type point of view : absolute
From the measurement scale point of view: interval
Type
From the calculus point of view: primary
From the time evolution point of view: moment
Class/ category Indicators of level and structure
Local, by county, regional, national, continental, worldwide, by gender, by
Aggregation level
residence environments, by ethnicity, etc.
If the minimum limit ( ε ) can be assured by 2-3 confessions a church – cartel
Interpretation can become the representatives in the relations between State and Church. If
OLIG=1 than we are in the situation of monopoly.
The indicator characterizes the religious market in a certain location. Even if at
a national or regional level the situation is in a certain way (one confession
Quality standards
dominates the market) there can be other locations where another confession
dominates the religious market, locally.

Indicator 4. The structure of the population from the viewpoint of active

Characteristics affiliation
The weights that religious institutions have within a population active from a
Definition
religious point of view.
It is calculated from the balance of religious institutions and the active
Scope
religious population, taking into consideration the IRI5.
Symbol ACTRLG
It is calculated by dividing the total number of persons affiliated to a religious
confession to the total investigated population. Facultative, the result can be
Calculus method ecpressed as percentage, by multiplication with 100. The declaration of a
person as belonging to a category of active believers s accomplished by using
the AB6 indicator (presented in the human resources indicators section).
ABi
ACTRLGi = ⋅ [100] (4.)
∑ ABi
Formula Where:
ABi – the number of active believers belonging to confession ranking i.
∑ AB i - The volume of the population of active believers.
Necessary data AB indicator– active believers.
Data source Statistical investigations, Religious institutions, NIS.
From the data type point of view : relative
Type From the measurement scale point of view: continuous
From the calculus point of view: derived

528
 Quantitative Methods Inquires

From the time evolution point of view: moment

Class/ category Indicators of level and structure
Local, by county, regional, national, continental, worldwide, by gender, by
Aggregation level
residence environments, etc.
Measurements and comparisons can be made between the existing situation
regarding active religious affiliation in various geographical locations, or in
Interpretation
population structured according to several aggregation criteria (gender,
education, nationality, etc).
The illustration precision of the existing situation for a certain location
increases if this indicator is utilized. Taking into consideration the
Quality standards secularization manifested in religion, sometimes there can be situations when
a religious minority (according to RLG) becomes the leader of a certain
religious market (using ACTRLG as study method).

Indicator 5. The balance of legally constituted families

Characteristics
The specific weight of families legally registered as against to the total number
Definition
of families.
The indicator shows, from a quantitative perspective, how the concept of
Scope
“family –basic cell of society” is perceived in the daily practice.
Symbol LEGFAM
The total number of legally constituted families is divided to the total number
Calculus method of families. The result can also be expressed as percentage, by multiplication
with 100.
LF
LEGFAM = ⋅ [100] (5.)
TNF
Formula
Where:
LF – the total number of legally constituted families.
TNF – the total number of families.
The number of legally constituted families, as well as the total number of
Necessary data
families.
Data source Statistical investigation, NIS, censuses.
From the data type point of view : relative
From the measurement scale point of view: continuous
Type
From the calculus point of view: derived
From the time evolution point of view: moment
Class/ category Indicators of level and structure
Local, by county, regional, national, continental, worldwide, by confessions, by
Aggregation level
residence environments, etc.
The more the indicator is closer to 1 (or 100%) the more the studied
Interpretation
population has the tendency to register legally constituted families.
Various interpretations can be given to the way of registration. Most of the
times, religious institutions don’t recognize a family unless it has also been
Quality standards
registered by a religious institution. There are also situations in which legal
registration is sufficient for religious registration7.

Indicator 6. The balance of multi-confessional families

Characteristics
The number of families that consist of at least two believers of another
Definition
religious institution compared to the total number of families.
It intends to illustrate the degree of opening, freedom, and overcome of
Scope
confessional barriers within a family.
Symbol MCONFAM
The number of multi-confessional families is divided to the total number of
Calculus method families. The result can also be expressed as percentage by multiplication with
100.

529
 Quantitative Methods Inquires

MCF
MCONFAM = ⋅ [1000] (6.)
TNF
Formula
Where:
MCF – the number of multi-confessional families
TNF – the total number of families
The number of multi-confessional families, as well as the total number of
Necessary data
families.
Data source Statistical investigations, NIS
From the data type point of view : relative
From the measurement scale point of view: continuous
Type
From the calculus point of view: derived
From the time evolution point of view: moment
Class/ category Indicators of level and structure
Local, by county, regional, national, continental, worldwide, by residence
Aggregation level
environments, by confessions, by ethnic groups, etc.
The interpretation can be performed by comparison to other countries or by
Interpretation
study of confessions where the multi-confessional families are more frequent.
Religious institutions usually tolerate but don’t encourage multi-confessional
families. This is the reason why the indicator will not register a high level. The
Quality standards
analysis can be refined by using the information regarding the religion chosen
by the children of the multi-confessional families.

Conclusions and future work

Without a doubt, the description within a standardized system and the creation of
a unitary and comprehensive conceptual context is merely the methodological foundation
necessary for the following phases: i) implementing improvements (after receiving feed-back
from the interested academic community), ii) collecting statistic information for the
alimentation and operationalization of this system, iii) analyzing, processing and
disseminating the results towards the scientific community and interested public.
We hope that this methodological process will be followed by the implementation
of the other phases necessary to finalize the scientific process initiated by this material.

References

1. Frunza, S. and Frunza, M Etica, superstitie si laicizarea spatiului public, Journal for the Study
of Religions and Ideologies, nr. 23, Spring 2009, pp. 13-35
2. Herteliu, C Metode cantitative in studiul fenomenului religios, Ed. Napoca Star, Cluj Napoca,
2009
3. Herteliu, C. Statistical Indicators System regarding Religious Phenomena, Journal for the
Study of Religions and Ideologies, no. 16, Spring 2007, pp. 111-127
4. Iannaccone, L., Introduction to the Economics of Religion, Journal of Economic Literature, Vol.
36, No. 3, September 1998, pp. 1465-1495
5. Idel, M. Abordari metodologice in studiile religioase, Journal for the Study of Religions and
Ideologies, volume 6 (16), 2007, pp. 5-20
6. Isaic-Maniu, Al. (coord.), Pecican, E., Stefănescu, D., Voda, V., Wagner, P. Dictionar de Statistica
Generala, Ed. Economica, Bucharest, 2003
7. Isaic-Maniu, Al. and Herteliu, C. Ethnic and Religious Groups in Romania – Educational
(Co)Incidences, Journal for the Study of Religions and Ideologies, No. 12, Winter
2005, pp. 68-75

530
 Quantitative Methods Inquires

8. Jonhnstone, J.N. Indicators, Educational. In: T.Husén, T.N. Postlethwaite (editors) The
International Encyclopedia of Education, Pergamon Press, Oxford, vol. V, 1985, p.
2432-2438
9. Smidt, C.E., Kellstedt, L.A. and Guth, J.L. (eds.) The Oxford Handbook of Religion and
American Politics, Oxford University Press, New Yourk, 2009, pp. 3-42

1
Acknowledgements
This paper was designed with support from the following research project: „The Influence of Religion on the Social
Processes from Romania– A Quantitative New-Weberian Approach Using Statistical Samples Techniques”, code
1809, Project Manager Claudiu Herteliu. Grant won in 2008 “Ideas” Competition from PNCD II and financial
supported by UEFISCSU.

2
Claudiu Herteliu currently holds the position of Assistant Professor within the Chair of Statistics and Econometrics,
at the Bucharest University of Economics. In 2007, he earned his PhD degree at the same university. His main areas
of academic interest are: theoretical and macroeconomic statistics, religion statistics, education statistics,
econometrics. He attended many scientific conferences both in Romania and abroad (USA, Germany, Holland,
France, Italy, Hungary etc.) He published (as co-author and single author) over 20 articles in scientific journals, and
contributed to a number of four books published in Romania.

3
National Institute of Statistics

4
In Romania, in the communist period, the censuses in 1956, 1966 and 1977 have no record of religious
affiliations, which makes time comparisons difficult.

5
It is about the Intensity of Religious Implication indicator.

6
It is about Active Believer indicator.

7
The case of the USA procedures, where in many states, the marriage license is given after the religious ceremony.

531
 Book Review

Valerica MARES
Ph.D, Lecturer, Faculty of Accounting and Management Information Systems,
University of Economics, Bucharest, Romania

E-mail: [email protected] Web page: www.cig.ase.ro

Key words: information systems audit concepts; corporative governance; systems

development life cycle; information system infrastructure and security; encrypt systems;
business continuity and disaster recovery

Book Review on
AUDIT AND INFORMATION SYSTEMS CONTROL
(“AUDITUL SI CONTROLUL SISTEMELOR INFORMATIONALE”),

by Pavel NASTASE (coord.), Eden ALI, Floarea NASTASE, Victoria

STANCIU, Gheorghe POPESCU, Mirela GHEORGHE, Delia BABEANU,
Dana BOLDEANU, Alexandru GAVRILA,
Ed. Economica, Bucharest, 2007

The necessity to adapt the audit to

the new technological reality, having in the
first plan the advantages generate by these
technologies and the risks of the
environments, has double impact: the
information technology is subject and object
of the audit.
The authors propose to identify the
vulnerabilities regarding the information
systems security and to elaborate different
forms to eliminate the threats.
The first chapter presents standards
and guides for information systems audit. It
is also present the planning, management
and deployment of the information systems
audit processes.
Chapter three is dedicated to the
information system life cycle management
and application controls. There are define
the life cycle, projects management,
development methods, including systems

532
 Book Review

development life cycle, risks associated, Business Process Reengineering (BPR).

Technical infrastructure of the information systems, chapter four, contain: hardware
and software architecture, operating in information systems, data management and network
infrastructure.
Chapter five has information systems security basic elements as security
management, logical access control, Internet and Intranet security, intruders’ detection
systems, security systems and electronic equipments protection.
Encrypt systems are the most important in data protection. In this chapter there are
presented encryption with secret and public key, digital signature, infrastructure of the public
keys and how we use the encryption in OSI protocols.
Business Continuity and Disaster Recovery are also important. In chapter seven,
there are presented the BCP components, information systems continuity and planning for
disaster recovery.
All seven chapters are finalized by questionnaires to determine the knowledge
gathered regarding the problems presented.
The paper work is necessary to the auditors, specialists in information systems
security, managers, but not the last to those who want to prepare CISA exam.
The preface of this book is presented by Alan T Lord, Professor of Accounting,
Bowling Green State University, Ohio USA, and he recommends:
“As the Chair of the global task force from ISACA, I think the authors have done a
commendable job with this text. The organizational structure and content of this text allow it to
serve as a preference primer for studying the information systems audit and control discipline.”

533