<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by BearingPoint Data, Analytics &amp; AI on Medium]]></title>
        <description><![CDATA[Stories by BearingPoint Data, Analytics &amp; AI on Medium]]></description>
        <link>https://medium.com/@bearingpoint-analytics-ai?source=rss-23175f7f65bd------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*E6O1TLZ6ydbCNOsIgGL74Q.png</url>
            <title>Stories by BearingPoint Data, Analytics &amp;amp; AI on Medium</title>
            <link>https://medium.com/@bearingpoint-analytics-ai?source=rss-23175f7f65bd------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 17:10:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@bearingpoint-analytics-ai/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Data-Driven Maintenance: Tailored insights for drastic improvements in operational efficiency]]></title>
            <link>https://bearingpoint-analytics-ai.medium.com/data-driven-maintenance-tailored-insights-for-drastic-improvements-in-operational-efficiency-da0248b224d9?source=rss-23175f7f65bd------2</link>
            <guid isPermaLink="false">https://medium.com/p/da0248b224d9</guid>
            <category><![CDATA[data-driven]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[iot]]></category>
            <dc:creator><![CDATA[BearingPoint Data, Analytics & AI]]></dc:creator>
            <pubDate>Tue, 15 Oct 2024 07:13:04 GMT</pubDate>
            <atom:updated>2024-10-15T08:30:23.101Z</atom:updated>
            <content:encoded><![CDATA[<h4>By August Emil Stokkeland (Senior Manager) and Vebjørn Axelsen (Partner) at BearingPoint</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h90Qb3VtcomCvWz2Cvi5mA.jpeg" /></figure><h4>Data-driven maintenance is a key component of Industry 4.0, utilizing technologies such as IoT, sensors, statistical analysis and machine learning to achieve the ultimate goal of reduced maintenance costs and shorter downtimes. But how can we effectively leverage these technologies to create lasting improvements in maintenance? Let’s explore the best approaches to optimize maintenance with a data-driven approach.</h4><p>At BearingPoint, we have extensive experience in guiding clients toward effective data-driven and predictive maintenance practices. In this article, we will describe the core elements of our methodology for data-driven maintenance improvements;</p><ul><li>Be smart in using data for effective, if not always predictive, maintenance</li><li>Let value, not technology, be the driving force behind realizing data-driven improvements</li><li>Work iteratively with rapid prototyping in cross-functional teams</li><li>Ensure that data-driven work methods are fully incorporated into operative maintenance</li></ul><h3>Be smart in using data for effective, if not always predictive, maintenance</h3><p>There are many discussions around predictive maintenance, and the potential benefits of implementing predictive models in maintenance processes are obviously big. However, predictive maintenance should not necessarily be a goal in and of itself. Sometimes it can be more effective to detect faults as they occur, using simpler models which analyze sensor data (often in real-time), based on an understanding of the workings of each component. Depending on the domain there may also be types of faults and failures that are near impossible to detect and correct before they happen, and where predictive maintenance will therefore not be useful.</p><p>Generally, we still wish to move from corrective to preventative, from the calendar based to the more advanced, data-intensive modes, as shown in the figure below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*27dktHk-IB6N27_E" /></figure><p>Data-driven maintenance is therefore about using data smartly, meaning cost-effectively and fit for purpose given the prioritized strategic goals to:</p><ul><li>Enable more advanced and data-intensive maintenance strategies where these are most suitable (condition-based, predictive, and prescriptive; where the latter both predicts the fault and recommends the action necessary to correct it).</li><li>Choose the most effective maintenance strategy for each component and type of fault or failure, based on available insight into what is most cost-effective in each case.</li><li>Effectively prioritize and control the actual maintenance work being performed day-to-day across components, fault/failure types, and maintenance strategies.</li></ul><p>It thus becomes clear that data-driven maintenance is about using data broadly and systematically, well beyond understanding the current and future condition of each component under maintenance.</p><h3>Let value, not technology, be the driving force behind realizing data-driven improvements</h3><p>To succeed with data-driven improvements we suggest some central principles, which build upon traditional capabilities such as change management and benefits realization during implementation.</p><h4>1. Be clear on strategic focus and potential</h4><p>Possible improvements must be identified based on concrete business goals, typically reduced costs and downtimes, so that we focus on what is strategically important and relevant.</p><h4>2. Identify concrete actions</h4><p>We must be completely clear on how actions and decisions in each process will be supported by targeted insight from data to create the necessary improvements.</p><h4>3. Think end-to-end</h4><p>The implementation of a data-driven improvement must be pushed all the way out into day-to-day processes in the maintenance organization. The pitfall here is to be content when the data/technical disciplines are under control, but that is of course not enough. Effective and continuous use of new insight from data demands changes in ways of working across systems and processes, as well as culture. This requires focus and effort over time to do right.</p><h4>4. Measure actual improvement effects</h4><p>We must measure the effects of the process changes we make in a structured way to verify that we create real value (such as reduced costs and downtimes). This kind of effect measurement is of course done with data, and our concept of <em>data-driven</em> thus applies not only to the underlying data analysis to understand the component conditions, but equally to inform how we should design and implement a process change.</p><p>In practice, this means designing from right to left in the figure below. Only when we have a good understanding of what type of insight from data is to be used where and for what, can the technological aspects around sensors, data and analysis be considered effectively.</p><figure><img alt="Figure: BearingPoint’s intangible rights" src="https://cdn-images-1.medium.com/max/1024/0*EhCoM6knUh9UYd6J" /><figcaption><em>Figure: BearingPoint’s intangible rights</em></figcaption></figure><p>Going back to the topic of designing based on business value: Data-driven maintenance will often influence the organization broadly, since there can be great value in using relevant data well beyond the most obvious operational use cases in condition-based and predictive maintenance. The figure below shows typical use case areas from the most operational to the more strategical, all of which find support in tailored insight from data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bHqXI6VjWg3mLUaN" /></figure><p>In the following paragraphs, we’ll focus on the most common place to start implementing data-driven maintenance: operational use cases in discovering and predicting faults and failures on equipment.</p><h3>Work iteratively with rapid prototyping in cross-functional teams</h3><p>To be successful it is important to not just focus on what is expected to be the most valuable improvement candidate (1 in the below figure), but also to confirm/reject candidates through rapid prototyping (2) so that organizational, data-related, and technical risks are addressed before spending time and resources in scaling implementation in the organization (3). This last part also encompasses the development of production ready technical solutions (sensors, data flows, analytical models, and integration with other systems on a scalable platform).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*W-uIawAYR4gaI9tn" /></figure><p>Through all three phases a close day-to-day collaboration between business functions (result owners, operators, and domain experts) and data science functions (data and model developers) is essential:</p><h4>Phase 1</h4><p>The most promising improvements are identified through discussions between the business and data science functions, where an in-depth understanding of the problem area (for example, specific knowledge about the equipment and reasons for failure) is combined with analytical creativity and realism (what data is available, and which analytical techniques can be applied to create the needed insight based on these data). Additionally, deep-dives into historical fault and failure data is useful to solidify business case estimates for improvement candidates in focus.</p><h4>Phase 2</h4><p>To accurately describe e.g. failure modes in analytical models we need very close (daily) collaboration between a domain expert and a data scientist, working together in short iterations to attack the problem from different angles, and jointly arrive at good solutions. If one had “perfect data”, where sensor data, historical failure data, and related master data was available completely and consistently in high volumes, machine learning algorithms could infer patterns for us quite easily. In practice, however, the data seldom meets such quality levels, and good collaboration between the disciplines is then essential to create relevant and reliable models using the data we have best-effort.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eypia99l8DjoXuZE" /></figure><h4>Phase 3</h4><p>Before roll-out, insight from the prototype phase should be used both to update the business case and to tailor the implementation. For example, when introducing a predictive maintenance mode for a certain component type, data regarding actual failure rates and the known precision of our failure prediction models should be the foundation for redesigning maintenance routines, so that the actions are planned optimally and cost/capacity for maintenance work is balanced against the risk tolerance of equipment downtime.</p><h3>Ensure that data-driven work methods are fully incorporated into operative maintenance</h3><p>To succeed with lasting and significant improvements, the data-driven ways of working and related data and models must be incorporated as a natural and fully integrated part of the maintenance system. This allows two success factors to be met:</p><ul><li><strong>Specific and tailored information integrated into the operative maintenance</strong><br>The creation of work orders (WO) based on recommendations from analytical models can be standardized. For equipment and failure types where we achieve high model precision in detecting/predicting faults, and where we know what action must be taken, the creation of work orders can be automized.</li><li><strong>Structured feedback from performed maintenance</strong><br>Ongoing, continuous feedback on an adequate scale (Was the alarm relevant, and what happened as a result?) can be used systematically, so that we can develop better analytical models. With increasing maturity and standardization in this process, feedback data can be used automatically for machine learning: Models that detect/predict faults and failures will continually learn and improve from feedback given by maintenance workers in the field.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bS2Vp5moYuVdqjBR" /></figure><p>This demands not only system technical integrations between analysis platforms and ERP/work order systems, but also a redesign of operational maintenance routines. With the latter comes also a change in culture, from strategic planners and all the way out to operative maintenance personnel in the field. A change like this can be demanding and take time, but with close collaboration across teams and a clear path to profit realization, the framework is in place to take gradual steps and to build experience, commitment, and maturity over time.</p><h3>Want to learn more? Get in touch!</h3><p>In conclusion, data-driven maintenance is not really about the data you collect. It is about understanding how maintenance processes can be improved by tailored use of data, strategically and operationally, focusing on the potential for significant and lasting value creation. At BearingPoint, we combine solid expertise in data and data science with broad experience in operational excellence across technology and business disciplines.</p><p>Get in touch to learn how we can help you use data smartly to increase uptime and reduce maintenance costs! Feel free to reach out to <a href="https://www.linkedin.com/in/august-emil-stokkeland-01b09630/">August Emil Stokkeland</a> or <a href="https://www.linkedin.com/in/vebjornaxelsen/">Vebjørn Axelsen</a> at BearingPoint.</p><h4>Want to find out more? <a href="https://www.bearingpoint.com/en-no/our-expertise/capabilities/technology/data-analytics/">Visit our website.</a></h4><h4>Find us on <a href="https://www.facebook.com/BearingPointNorway/?ref=bookmarks">Facebook</a> and <a href="https://www.instagram.com/bearingpoint_norge/">Instagram</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da0248b224d9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Transformers for Time-Series Data]]></title>
            <link>https://medium.com/bearingpoint-data-analytics-ai/transformers-for-time-series-data-3fadff9f07d8?source=rss-23175f7f65bd------2</link>
            <guid isPermaLink="false">https://medium.com/p/3fadff9f07d8</guid>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-and-analytics]]></category>
            <category><![CDATA[transformers]]></category>
            <dc:creator><![CDATA[BearingPoint Data, Analytics & AI]]></dc:creator>
            <pubDate>Tue, 19 Dec 2023 10:50:33 GMT</pubDate>
            <atom:updated>2023-12-19T13:57:10.555Z</atom:updated>
            <content:encoded><![CDATA[<h4>By Lars Ødegaard Bentsen (Data Scientist) at BearingPoint</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*g5PzWOlKYrmbD-7muzx3jQ.jpeg" /></figure><h4><strong>Since its release in 2017, the Transformer has brought with it a mini-revolution in the field of AI. By alleviating the need for recurrence, the Transformer was found to be able to learn longer contexts than traditional LSTM/RNN-based architectures.</strong></h4><p>Nevertheless, the Transformer did come with some limitations. Since the attention mechanism performs attention over all inputs, the model’s complexity scales quadratically with sequence length L². Other sequence learners, such as LSTMs, do not possess this drawback, with memory complexity that scales linearly with sequence length (O(L)). The quadratic memory growth of the Transformer might restrict the maximum sequence length of the inputs, depending on the readily available compute resources.</p><p>Furthermore, since the original Transformer (and much of the related work) focuses on natural language and computer vision applications, the original architecture does not necessarily work well on time-series data out-of-the-box.</p><p>In this article, we aim to provide a brief overview of some interesting work that adapts the Transformer architecture to facilitate time series data (and in particular forecasting) or that aims to reduce the model complexity (&lt;O(L^2)). Since only a brief overview of the different architectures will be provided here, we suggest that the interested reader checks out the full papers as well. Happy reading, and remember, <em>Attention is All you Need</em>!</p><h4><strong>Attention mechanism and Transformer</strong></h4><p>Attention mechanisms were first introduced for machine translation, where the attention operation enabled a recurrent encoder-decoder architecture to retain long-context information better than architectures without attention. The nice thing about attention mechanisms is that they’re fairly intuitive and easy to understand. As an example, we could visualise the learned attention weights for an image-captioning model, to see which parts of the image the model learns to focus on when producing a particular word.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/903/1*fwJ4HZdVZ8kwYEAjmoRfoQ.png" /><figcaption>Figure 1: Attention weight illustration for image captioning. Taken from: Xu, Kelvin, et al. “Show, attend and tell: Neural image caption generation with visual attention.” <em>International conference on machine learning</em>. PMLR, 2015.</figcaption></figure><p>From the image, we can see that the attention mechanism learns to focus on the correct parts of the image for different words (e.g. focusing on the central parts when producing the words bird and flying). Pretty cool!</p><p>Moving on to the Transformer architecture, the author’s idea was to design an architecture fully reliant on the attention mechanism, rather than combining attention with conventional convolutional or recurrent architectures. The overall architecture is visualised in the image below, along with an illustration showing how the scaled-dot-product attention works.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/461/1*kKVGFsVDfQwGw9sJMyYdiQ.png" /><figcaption>Figure 2: The Transformer Architecture from: Vaswani, Ashish, et al. “Attention is all you need.” <em>Advances in neural information processing systems</em> 30 (2017).</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_92bnsMJy8Bl539G4v93yg.gif" /><figcaption>Figure 3: Illustration of the scaled dot-product attention. This great illustration is taken from another blog post: <a href="https://towardsdatascience.com/illustrated-self-attention-2d627e33b20a">https://towardsdatascience.com/illustrated-self-attention-2d627e33b20a</a></figcaption></figure><p>We’re not going to dig deep into the Transformer architecture here. The interested reader may however check out the references at the end of this post or the Illustrated Transformer article here: <a href="https://towardsdatascience.com/illustrated-self-attention-2d627e33b20a">https://towardsdatascience.com/illustrated-self-attention-2d627e33b20a</a></p><h4><strong>Some Prominent Architectures</strong></h4><p>If we were to visualise the attention pattern for the vanilla (“basic/original” in deep learning lingo) Transformer, we would require an LxL matrix to store all the attention weights. This is illustrated in the top-left of the figure below and results in quadratic memory growth.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/246/1*2icMsMi2h4wqtoqq4b5ahw.png" /><figcaption>Figure 4: Illustration of different attention patterns.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/903/1*EKG7ZgXKkj5ViqUWUtuzlw.png" /><figcaption>Figure 5: Another visualisation that shows the difference between windowed (right) and full (left) attention for an NLP application with a sentence input.</figcaption></figure><p>The first few architectures that we will look at, focus on alleviating the challenge of quadratic memory growth by introducing sparsity into the attention operation. This means that keys will no longer be able to attend to all other sequence positions in an update (as seen by the three other attention patterns in the figure). We will now provide a summary of the key features of some prominent Transformer architectures designed for time series or long-sequence data:</p><p><strong>Longformer (referring to figure 4)</strong></p><ul><li>This alteration of the Transformer was designed to be able to work on very long documents, which prohibits the use of full self-attention (due to memory constraints).</li><li>Diluted windowed attention: Each sequence position can only attend to nearby positions, as shown in the bottom left of the figure above.</li><li>To learn long-term context, the authors also introduced global attention, where certain sequence positions can be attended to by all other keys (see the full paper for details).</li></ul><p><strong>LogSparse Transformer (referring to figure 6)</strong></p><ul><li>Convolutional Attention: The full self-attention of the Transformer uses position-wise linear transforms to compute the key/query representations, which are insensitive to local context. By leveraging 1D convolution, the key/query representations can now contain local information, such as mean/std for the local time interval or be able to identify increasing/decreasing trends. This might be very useful for time-series applications where you have high volatility and more chaotic signals than for natural language. One isolated time step may not provide as much information as a single word does within a sentence, due to the (typically) noisy nature of time series signals. This is a motivation behind employing convolutional attention for analysing time series data, as it can help to capture local context and relationships between different time steps.</li><li>LogSparse Attention: Some of the memory constraints are alleviated by introducing sparsity into the attention patterns (as for the Longformer) based on an exponential step schedule. The authors also introduce some local attention range and restart (as visualised in the figure below).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/602/1*A73MwHunYUAhgpVS7nYIAQ.png" /><figcaption>Figure 6: Illustration of full and LogSparse attention patterns on time series data. Taken from: Li, Shiyang, et al. “Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting.” <em>Advances in neural information processing systems</em> 32 (2019).</figcaption></figure><p><strong>Informer (refering to figure 7)</strong></p><ul><li>Sparsity: Instead of having sparse attention patterns that are decided based on (simple) fixed heuristics, the Informer introduces the ProbSparse attention. The ProbSparse attention locates the most dominant queries and only allows keys to attend to these (based on the Kullback-Leilbler (KL) divergence against a uniform distribution). In essence, this method of introducing sparsity should not overlook potentially important time steps, which might be the case for fixed sparsity patterns.</li><li>Self-attention distilling: Highlights dominant attention by halving cascading layer inputs through 1D convolution and max pooling. If this is difficult to understand, the distilling makes the Informer much more efficient for very long sequences.</li></ul><p><strong>Autoformer (refering to figure 7 and 8)</strong></p><ul><li>Auto-Correlation: The Autoformer replaces the original scaled dot-product attention mechanism with the Auto-Correlation module (!). This operation uses keys and queries to decide on the most important time-delay similarities through autocorrelation and time-delay aggregation. This again, is expected to perform better for many time series applications, as it is specifically designed for time series similarity.</li><li>Signal decomposition: The Autoformer also fundamentally changes the vanilla Transformer architecture. We will not dive into the details here, but a key takeaway is the use of signal decomposition. Series decomposition is used to separate trend and seasonal components (important!) in the input signal and the Autoformer then performs computation on periodic (seasonal) signals to extrapolate into the future.</li><li>Overall, the Autoformer presented a shift in Transformers for time-series analysis, as it directly changed the Transformer architecture to facilitate time-series data, based on established concepts within signal processing.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/602/1*4BM6taJc2WwTKnomeDdSrA.png" /><figcaption>Figure 7: Attention patterns for (a) — Vanilla Transformer, (b) — any sparse attention, e.g. ProbSparse from Informer, © — LogSparse Transformer attention, (d) — Auto-Correlation from Autoformer. Taken from: Wu, Haixu, et al. “Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting.” <em>Advances in Neural Information Processing Systems</em> 34 (2021): 22419–22430.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/451/1*4L3FUmLyDzKQREoXLPKQqA.png" /><figcaption>Figure 8: The Autoformer Architecture. Taken from: Wu, Haixu, et al. “Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting.” <em>Advances in Neural Information Processing Systems</em> 34 (2021): 22419–22430.</figcaption></figure><p><strong>FFTransformer (refering to figure 9)</strong></p><ul><li>Decomposition and separate streams: The architecture uses wavelet decomposition to partition the original time series into trend and periodic components. The architecture then performs computation on the two types of signals, separately.</li><li>Trend Stream: Computation on trend (<strong>blue</strong>) signals is similar to the original Transformer.</li><li>Periodic stream: Fast Fourier transform (FFT) is applied to the periodic (<strong>green</strong>) signals before an attention mechanism updates signals in the frequency domain.</li><li>Output: Latent representations from the two streams are combined and passed through a linear layer to produce the final outputs.</li><li>Sparsity: Although the FFTransformer does not introduce any new methods for reducing the memory complexity, the architecture allows for any attention operation to be used (e.g. ProbSparse, LogSparse, windowed attention).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/602/1*wps2EalafqJjzdFEu3mtqg.png" /><figcaption>Figure 9: Example of decomposing a signal (black) into periodic (green) and trend (blue) components. Taken from: Bentsen, Lars Ødegaard, et al. “Spatio-temporal wind speed forecasting using graph networks and novel Transformer architectures.” <em>Applied Energy</em> 333 (2023): 120565.</figcaption></figure><h4><strong>Key takeaways</strong></h4><p>You have now learned some key concepts related to the Transformer architecture and things that you might consider when aiming to use Transformers for time-series data (or maybe other applications?). As a few final remarks, we can summarise what we’ve learned:</p><p><strong>Introducing sparsity into the attention mechanism can help alleviate memory constraints</strong>. Sparsity could be based on either fixed heuristics (LogSparse, Windowed, Global+Diluted Windowed Attention) or by locating the dominant attention pairs (ProbSparse Attention, Auto-Correlation).</p><p><strong>Time series are often fundamentally different from text-based data and one can potentially see huge improvements by leveraging traditional concepts from signal theory </strong>(Don’t forget your good old maths/engineering concepts!). For example, by decomposing a signal into trend and periodic components, it can be easier for a model to learn overall characteristics and not be confused by high volatility/noisy signals.</p><p><strong>Specialized attention operations that operate in the frequency domain (Autoformer, FFTransformer, FEDformer), can help boost performance.</strong> These, as well as other methods such as Convolutional Attention, can help provide local context to the model. Such alterations might be important for time-series signals, where recordings are not necessarily very informative in isolation.</p><p><strong>Get to know your time series data</strong> and see if there are any traditional concepts (e.g. signal decomposition) that can be used by the model to potentially boost performance! Furthermore, it is always good to do a sanity check before applying any deep learning model to see if it makes sense for the model to learn the desired characteristics, e.g. convolutional attention (from LogSparse Transformer) can help with local context.</p><h4><strong>About the author</strong></h4><p><a href="https://www.linkedin.com/in/bentsenlars/"><strong>Lars Ødegaard Bentsen</strong></a> is a Consultant in the Data Science and AI team at BearingPoint Oslo. He recently handed in his PhD in DL applied to offshore wind energy, where he mainly focused on spatio-temporal time series forecasting. Email address: <a href="/service/mailto:lars.bentsen@bearingpoint.com">lars.bentsen@bearingpoint.com</a></p><p><strong>Side Note/Disclaimer:</strong> You might also check out some of the references for additional architectures that we excluded from this blog post for brevity. The field is also moving quite rapidly, so do not use this post as an exhaustive list of architectures, but rather as a soft guide into Transformers for time series!</p><h4>References</h4><ul><li><strong>Original Attention Paper: </strong>Bahdanau, Dzmitry, Kyunghyun Cho, and Yoshua Bengio. “Neural machine translation by jointly learning to align and translate.” <em>arXiv preprint arXiv:1409.0473</em> (2014).</li><li><strong>Image Captioning with Attenion:</strong> Xu, Kelvin, et al. “Show, attend and tell: Neural image caption generation with visual attention.” <em>International conference on machine learning</em>. PMLR, 2015.</li><li><strong>Transformer Paper:</strong> Vaswani, Ashish, et al. “Attention is all you need.” <em>Advances in neural information processing systems</em> 30 (2017).</li><li><strong>Longformer Paper:</strong> Beltagy, Iz, Matthew E. Peters, and Arman Cohan. “Longformer: The long-document transformer.” <em>arXiv preprint arXiv:2004.05150</em> (2020).</li><li><strong>LogSparse Paper:</strong> Li, Shiyang, et al. “Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting.” <em>Advances in neural information processing systems</em> 32 (2019).</li><li><strong>Informer Paper:</strong> Zhou, Haoyi, et al. “Informer: Beyond efficient transformer for long sequence time-series forecasting.” <em>Proceedings of the AAAI conference on artificial intelligence</em>. Vol. 35. №12. 2021.</li><li><strong>Autoformer Paper:</strong> Wu, Haixu, et al. “Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting.” <em>Advances in Neural Information Processing Systems</em> 34 (2021): 22419–22430.</li><li><strong>FFTransformer Paper:</strong> Bentsen, Lars Ødegaard, et al. “Spatio-temporal wind speed forecasting using graph networks and novel Transformer architectures.” <em>Applied Energy</em> 333 (2023): 120565.</li><li><strong>Reformer Paper:</strong> Kitaev, Nikita, Łukasz Kaiser, and Anselm Levskaya. “Reformer: The efficient transformer.” <em>arXiv preprint arXiv:2001.04451</em> (2020).</li><li><strong>FedFormer Paper:</strong> Zhou, Tian, et al. “Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting.” <em>International Conference on Machine Learning</em>. PMLR, 2022.</li></ul><h4>Want to find out more? <a href="https://www.bearingpoint.com/en-no/our-expertise/capabilities/technology/data-analytics/">Visit our website.</a></h4><h4>Find us on <a href="https://www.facebook.com/BearingPointNorway/?ref=bookmarks">Facebook</a> and <a href="https://www.instagram.com/bearingpoint_norge/">Instagram</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3fadff9f07d8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bearingpoint-data-analytics-ai/transformers-for-time-series-data-3fadff9f07d8">Transformers for Time-Series Data</a> was originally published in <a href="https://medium.com/bearingpoint-data-analytics-ai">BearingPoint Data, Analytics &amp; AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Working with large amounts of data in Power BI]]></title>
            <link>https://medium.com/bearingpoint-data-analytics-ai/working-with-large-amounts-of-data-in-power-bi-4f3dea3906ee?source=rss-23175f7f65bd------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f3dea3906ee</guid>
            <category><![CDATA[data-insights]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[BearingPoint Data, Analytics & AI]]></dc:creator>
            <pubDate>Tue, 22 Aug 2023 08:23:10 GMT</pubDate>
            <atom:updated>2023-08-22T08:23:10.559Z</atom:updated>
            <content:encoded><![CDATA[<p>By Fredrik Bonnevie Dahler (Senior Technology Consultant), Ulrika Axelsson (Technology Analyst) and Mikkel Boye Ahlgren (Senior Technology Consultant) at BearingPoint</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G4vsCkyW1W9nu46XteZGGg.jpeg" /></figure><p><strong>Many companies have large amounts of data in their data warehouses. In order to deliver a good user experience, reports both need high performance and have the granularity to drill into the details of the data. However, the need for near-instant access to large amounts of data is in many cases difficult to align with the size limitations of in-memory Power BI Datasets. Delivering high-performance reports on large amounts of data is possible, but it requires a combination of good design choices and sound modeling principles.</strong></p><p>We will also delve into some aspects of a potential game-changer that came along with the recent introduction of Fabric, the Direct Lake mode for data access.</p><p>In this article, we highlight some important aspects to keep in mind when working with large amounts of data in Power BI:</p><ul><li>The Implications of using Import or DirectQuery, and the promises of Direct Lake mode.</li><li>Understanding how storage in Power BI works.</li><li>Rules of thumb for reducing the model size.</li></ul><blockquote><strong>Choosing between Import and DirectQuery is a trade-off between performance needs, modelling flexibility and required storage optimization efforts. Is Direct Lake the catalyst?</strong></blockquote><p>It is not uncommon that companies have transaction datasets with millions of rows of data per day, containing numerous columns, many of which might have high cardinality (a large set of unique values). One of the fundamental design choices for a Power BI dataset is whether to use DirectQuery or Import mode for the fact tables.</p><p>An aspect that makes this a hard decision is that neither Import nor DirectQuery are flawless solutions. Depending on the subscription type, using Import mode sets a limit on the amount of data we can use without significant pre-aggregation efforts. On the other hand, DirectQuery has no size limitations on the underlying source data but comes with drawbacks such as higher response time,<strong> </strong>limited modeling flexibility, and limits on the number of rows that can be returned in a query. Below are some considerations when choosing between the two options:</p><p><strong>DirectQuery avoids loading most data into Power BI but has limitations on performance and modelling flexibility</strong></p><p>Using fact tables with DirectQuery comes with several implications. Here is a list of some of the most important ones:</p><ul><li>Data is only loaded into Power BI when the report queries the underlying data source.</li><li>Limited data caching and each visualization object sends a query directly to the database.</li><li>Each interaction with the report (sorting, using slicers, etc.) triggers a new query.</li><li>Limits exist on the number of concurrent queries that can be sent and the number of rows that can be returned.</li><li>There exist modelling limitations in both DAX and PowerQuery when using DirectQuery (For further insight, look at this article <a href="https://radacad.com/directquery-connection-in-power-bi-how-does-it-work-limitations-and-advantages">https://radacad.com/directquery-connection-in-power-bi-how-does-it-work-limitations-and-advantages</a>).</li><li>High costs of querying to underlying data sources: Solutions such as BigQuery and Snowflake incur costs based on the amount of data queries. Depending on how the Power BI dataset is set up, the amount of data queried, and the number of users, a DirectQuery setup might incur high Data Warehouse costs.</li></ul><p>Benefits of DirectQuery:</p><ul><li>Avoids data loading and storage issues with large models.</li><li>Data in the report is immediately updated along with changes in the source.</li></ul><p>Drawbacks of DirectQuery:</p><ul><li>Report performance depends on the source query speed.</li><li>Report performance is reduced when visualizing large amounts of data.</li><li>With a high number of concurrent users, the user experience is affected if the dataset approaches its limit on number of generated queries.</li><li>Lower modelling flexibility. Some effects are longer lead time on insights, decreased autonomy of teams and analysts/data generalists, higher load on data teams, or more generally, more difficult to do data democratization.</li></ul><p>To increase performance we can make use of dynamic PowerQuery parameters (m-parameters). This allows filters to be applied at the innermost level of the source queries.</p><p>Aggregation tables can be created to make import-versions of frequently used queries. Despite being useful to improve model performance, this technique doesn’t work well for dynamic reports where users frequently drill down into highly granular data.</p><h4>Import mode has high performance and modelling flexibility, but has limitations on model size and data update time</h4><ul><li>There are limitations on the dataset size. Premium Per User has a 100GB size limit, and a P2 Premium Capacity has a 50GB size limit (as does S2 for Azure Analysis Services). In practice, the size limits are about half the original size since half of the capacity needs to be allocated during model refresh <strong>(</strong><a href="https://learn.microsoft.com/en-us/power-bi/enterprise/service-premium-large-models#considerations-and-limitations"><strong>https://learn.microsoft.com/en-us/power-bi/enterprise/service-premium-large-models#considerations-and-limitations</strong></a><strong>)</strong>. Workarounds may involve splitting the data into smaller datasets, but this may not be an ideal solution as it requires maintaining multiple datasets.</li><li>Importing data increases the lead time for new data. Refreshing an import dataset adds additional lead time to the process of getting new data to end users. In some cases, we have experienced that refresh times can take up to 2.5 hours. Partitioning can be a relevant tool for reducing refresh time when data volume is the main contributor to time consumption, but in some instances it turns out that it really is compressing columns and indexing that are primary drivers. This issue is avoided with DirectQuery.</li></ul><p>Let us suppose that we prefer modelling flexibility and performance, and accept that we need to optimize our Power BI datasets to cater for the Import mode storage limitations. What should we consider? There are a few techniques and tools that can be leveraged to do so. In the following sections, we will first explain the basics of how storage in Power BI Datasets works and later discuss techniques and tools for optimizing them.</p><h4><strong>Direct Lake promises both fast performance and live connections to parquet-files</strong></h4><p>While DirectQuery mode is slow but real-time, Import mode is fast but requires duplicating data, refreshes in order to update a dataset, as well as maintenance of Power BI datasets in general. Direct Lake mode promises to deliver the best of both worlds.</p><p>Direct Lake mode reads directly on parquet-files in OneLake (which is built on top of ADLS Gen 2). This removes the need for maintaining Power BI datasets, and enables report updates as soon as the parquet-file changes in OneLake. Furthermore, there is no translation to other query languages or execution in other database systems, which according to Microsoft will result in performance comparable to import mode.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l-iFV4zHg-qt4C12RuyRsw.png" /></figure><p>Source: Microsoft (<a href="https://learn.microsoft.com/en-us/power-bi/enterprise/directlake-overview">https://learn.microsoft.com/en-us/power-bi/enterprise/directlake-overview</a>)</p><p>Fabric and Direct Lake are still in public Preview at the time of writing. Furthermore, in order to make use of this feature, you need a Premium Per Capacity or Microsoft Fabric license. See details here: <a href="https://learn.microsoft.com/en-us/power-bi/enterprise/directlake-overview#prerequisites">https://learn.microsoft.com/en-us/power-bi/enterprise/directlake-overview#prerequisites</a>.</p><p>It remains to be seen whether Direct Lake can eliminate all the drawbacks of the two current options, or if there will be use cases where each mode has an advantage over the other. While Direct Lake is worth monitoring closely, the current fully available solution requires a balancing between Import and DirectQuery mode. In the next sections, we’ll dive deeper into how data is stored in Power BI and how to handle storage issues when the amount of data grows large.</p><h4><strong>Power BI Datasets optimize storage under the hood, but tools also exist to facilitate manual storage optimization</strong></h4><p>Understanding how data is stored in Power BI is somewhat opaque. Tuning the storage and performance of Power BI datasets requires specific knowledge of how Power BI works under the hood. This topic is often not covered in tutorials, certification curriculums, or the Power BI documentation. This makes it difficult for data analysts or engineers to know what to consider when tuning Power BI datasets for handling large amounts of data.</p><p>Power BI data is stored in a database called VertiPaq, which is a columnar database. VertiPaq provides several advantages. Each column has its own structure and is physically separated from other columns. This separation allows for effective compression techniques to be applied independently to each column. Depending on the values within a specific column, VertiPaq selects the most suitable compression algorithm to minimize storage requirements.</p><p>Another important aspect is that VertiPaq is an in-memory database. This means that data is loaded and processed directly in memory, resulting in faster data retrieval and analysis, but is also why import mode has storage limitations.</p><p>A great tool for optimizing Power BI Datasets is VertiPaq Analyzer (https://www.sqlbi.com/tools/vertipaq-analyzer/). This powerful tool provides in-depth analysis and optimization recommendations for Power BI data models. By examining the VertiPaq engine’s performance and data compression, it helps identify potential bottlenecks and suggests improvements to enhance query response times, reduce memory consumption, and optimize data model design for better overall performance.</p><p>There are several techniques for optimizing storage in Power BI datasets. An in-depth elaboration on approaches and considerations is described in this brilliant article: “How to reduce your Power BI model size by 90%!” (https://towardsdatascience.com/how-to-reduce-your-power-bi model-size-by-90–76d7c4377f2d.) The sections below summarize the essence, and mentions relevant examples:</p><h4><strong>There are several ways to reduce the model size: Reducing the number of columns and cardinality are important levers</strong></h4><p>Here are some general rules for reducing the data model size:</p><ol><li>Include only the necessary columns and rows in the report.</li><li>Optimize column cardinality and evaluate the benefits of splitting columns.</li><li>Minimize the use of calculated columns, as they are not optimally compressed.</li><li>Use appropriate data types based on the required granularity.</li></ol><p><strong>Focusing on rules 1 and 2:</strong></p><ol><li>Import only the columns that are genuinely needed.</li></ol><ul><li>This reduces the file size and optimizes memory consumption.</li><li>In a significant example from our experience, we found a “Transaction-ID” column that wasn’t used in any of the visualizations on top of the dataset. The column contained a distinct string value for each transaction, and accounted for 40% of the total model size.</li></ul><p>2. Reduce column cardinality.</p><ul><li>Higher cardinality makes it harder for the data storage engine to compress the data optimally.</li><li>Split columns or modify data to reduce cardinality when possible. A typical example is splitting a Timestamp-column into Date and Time. Where a timestamp column over a year would have a maximum cardinality of 31 536 000 (365 * 24 * 60 * 60), splitting the column into separate Data and Time columns would reduce the cardinality to 86 765 (365 + 24 * 60 * 60). That’s a reduction of 98% (!).</li><li>Splitting integer columns or decimal values can result in significant savings.</li><li>Optimize date/time columns by removing unnecessary granularity if it meets the users’ requirements.</li><li>Aggregate data whenever possible to reduce cardinality and the number of rows.</li><li>Disable the Auto Date/Time option for data loading to eliminate unnecessary date tables (https://towardsdatascience.com/tiq-part-1-how-to-destroy-your-power-bi-model-with-auto-date-time-8fec32b22aff).</li></ul><p>In conclusion, managing large volumes of data in Power BI requires careful choices for both design and implementation. Import mode is attractive because it offers performance and modeling flexibility, but it necessitates optimizing the dataset size. To achieve this, the VertiPaq Analyzer tool and the provided rules of thumb can help ensure that your dataset remains within its size limitations.</p><h3>About the authors</h3><p><a href="https://www.linkedin.com/in/fredrik-bonnevie-dahler-649443170/"><strong>Fredrik Bonnevie Dahler</strong></a> is a Senior Technology Consultant at BearingPoint in Oslo, with experience in data and analytics from various industries including the financial industry, retail, and data products. He has focused on a business-centric approach to deliver insights for stakeholders. Before joining BearingPoint in 2019, Fredrik obtained a master’s degree in mechanical engineering from the Norwegian University of Science and Technology. Email address: <a href="http://fredrik.bonnevie.dahler@bearingpoint/">fredrik.bonnevie.dahler@bearingpoint</a>.</p><p><a href="https://www.linkedin.com/in/ulrika-axelsson-a0898a12a/"><strong>Ulrika Axelsson</strong></a> is a Technology Consultant in the Data &amp; Analytics team at BearingPoint in Oslo, with experience working with Power BI on several projects. Ulrika obtained a master’s degree in Industrial Engineering and Management from Umeå University in 2020. Email address: ulrika.axelsson@bearingpoint.com</p><p><a href="https://www.linkedin.com/in/mikkelahlgren/"><strong>Mikkel Ahlgren</strong></a> is a Senior Consultant in the Data &amp; Analytics team at BearingPoint in Oslo. He has broad experience in creating commercial value by leveraging data and analytics. He has a Master’s degree from NTNU in Industrial Economics and Technology Management, specialized in computer science and optimization, and a Bachelor’s degree from NHH in economics and business administration. Email address: <a href="/service/mailto:mikkel.ahlgren@bearingpoint.com">mikkel.ahlgren@bearingpoint.com</a></p><h3>Want to find out more? <a href="https://www.bearingpoint.com/en-no/our-expertise/capabilities/technology/data-analytics/">Visit our website.</a></h3><h3>Find us on <a href="https://www.facebook.com/BearingPointNorway/?ref=bookmarks">Facebook</a> and <a href="https://www.instagram.com/bearingpoint_norge/">Instagram</a>.</h3><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f3dea3906ee" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bearingpoint-data-analytics-ai/working-with-large-amounts-of-data-in-power-bi-4f3dea3906ee">Working with large amounts of data in Power BI</a> was originally published in <a href="https://medium.com/bearingpoint-data-analytics-ai">BearingPoint Data, Analytics &amp; AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A tailored Data Mesh for reMarkable]]></title>
            <link>https://medium.com/bearingpoint-data-analytics-ai/a-tailored-data-mesh-for-remarkable-e033e80118c6?source=rss-23175f7f65bd------2</link>
            <guid isPermaLink="false">https://medium.com/p/e033e80118c6</guid>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[data-management]]></category>
            <category><![CDATA[data-mesh]]></category>
            <category><![CDATA[data-insights]]></category>
            <category><![CDATA[data-infrastructure]]></category>
            <dc:creator><![CDATA[BearingPoint Data, Analytics & AI]]></dc:creator>
            <pubDate>Thu, 05 Jan 2023 09:46:38 GMT</pubDate>
            <atom:updated>2023-01-09T15:33:38.464Z</atom:updated>
            <content:encoded><![CDATA[<p>By Karan Kathuria (Head of AI Products) at BearingPoint and Martí Colominas (Head of Data &amp; Insight) and Frøydis Høistad (Data Governance Lead) at reMarkable</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5_9_FedSuiBzvloFUlxMbQ.jpeg" /></figure><h4>Data Mesh is a buzzword in the world of data management. At first glance, the core idea is inarguably very intriguing. On paper, Data Mesh offers to solve well known choke points related to an organization’s growing demand for data and analytics. But does it answer everyone’s big data challenges? Probably not.</h4><p>Few organizations are ready for a prompt transition into the purest version of the Data Mesh paradigm. However, we strongly believe that there are, depending upon the characteristics of your organization, many ideas within this paradigm that your organization could benefit from adopting. At reMarkable we’ve taken the core Data Mesh concept and developed a tailored hybrid version of Data Mesh.</p><h4><strong>Data Mesh in a nutshell</strong></h4><p>At its core, we regard Data Mesh as the idea of distributing the responsibility to develop and serve data as a product within an organization. Enabling this is far from easy, but by leveraging domain expertise this can be done in a much more agile manner than with a centralized responsibility. The key benefits of Data Mesh are arguably a scalable and flexible solution to the increasing demands for data and insight.</p><p>At <a href="https://remarkable.com/">reMarkable</a> we acknowledge that this transition can be challenging, but we do not regard the idea of decentralization as binary. We have evaluated the pros and cons related to both fully centralized and fully decentralized orientations in search of the right balance between the two.</p><h4><strong>The tailored Data Mesh</strong></h4><p>A typical data and analytics platform consist of many parts. The stack and infrastructure, the data content, the insights, the people, and the governance surrounding it, to mention some. Each of these parts could be placed somewhere on the axis between centralized and decentralized, making the question of centralized/decentralized, multi-dimensional.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pK4RPJq2LiGPVwfQEWaONQ.png" /></figure><p>The process of placing these parts on the axis required us to access some organizational attributes. The following attributes were among those considered:</p><ol><li>Ambitions related to the scale of data and analytics, both in terms of efficiency and size of the team</li><li>The current level of established common conventions and practices</li><li>Organizational culture, especially the level of autonomy within different business domains</li><li>The level of data literacy</li><li>The extent to which input data sources could be exclusively mapped to one business domain</li><li>The relative importance of domain expertise in materializing insight use cases</li></ol><p>In reMarkable’s case, the key drivers for a decentralized setup were:</p><ul><li>A strong culture for autonomy among and within the business domains. Each business domain sets its own prioritizations and has historically solved its insight requirements as they’ve found fit, with limited company-level guidance with regards to technical conventions, practices, and templates</li><li>High level of data literacy. reMarkable has a young and highly analytically oriented workforce. A large part of their workforce has a technical background and their ability to articulate insight requirements suggest a high level of data literacy</li><li>Level of ambition and future scale of data and insight. A shift in the business model suggested that the complexity and scale of insight requirements would increase rapidly. Enabling a data- and insight platform that would support this future scaling without creating future choke points was a key argument</li><li>Data domain and Business domain mapping. When trying to map out how each of the data domains (and related sources of data) within the organization would map to the business domain, we saw a surprisingly good match with few overlaps. This supported the idea of decentralizing data ownership making each domain responsible for its own data</li></ul><p>On the contrary, the following key driver argued for a more traditional centralized setup:</p><ul><li>The number of siloed analytics initiatives, the lacking technical conventions, naming conventions, templates, data interoperability and established best practices on how to build data pipelines and where to place business logic were all key arguments for starting out with a more centralized setup</li><li>A limited number of data engineers required us to be pragmatic in how we efficiently share a pool of data engineers leading us to centralize the company-wide prioritization of their efforts related to ingesting, refining, and managing data</li></ul><p>In order to balance these two forces, we came up with a compromise where we fully centralized the platform and technical infrastructure, and partly decentralized parts such as data governance, data ingest, and the pool of data engineers. In this, we found it highly useful to distinguish between source-aligned data — mostly meaning raw data from its source and consume-aligned data — meaning data as the business consumes it.</p><p>We’ve also ensured most data analysts can work end to end, becoming more analytical engineers than pure data analysts. By doing so, we could keep the engineering-intensive work related to ingesting, cleaning, and ensuring data interoperability more centralized, and the domain-intense work related to creating consumed aligned data products more decentralized. The thinking is that this would support a high level of autonomy and scalability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X7YJ3-rsm0nPq5LX8bpfWQ.png" /></figure><p>Overall, we ended up with a compromise placing us somewhere between a Data Mesh and a traditional centralized approach. We have done so with a clear intention of moving towards decentralization as we mature in terms of common conventions and practices and, increase our scale.</p><h4><strong>Leverage relevant ideas from the Data Mesh concept</strong></h4><p>BearingPoint supported reMarkable in establishing their target state for Data &amp; Insight, and was also central in establishing the organization and realizing the platform. In our view, organizations can and should benefit from leveraging relevant ideas from the Data Mesh concept, without having to adapt it completely. Feel free to <a href="/service/mailto:karan.kathuria@bearingpoint.com">contact us </a>for an informal chat on how this can improve your current state of data and analytics.</p><h4>Want to find out more? <a href="https://www.bearingpoint.com/en-no/our-expertise/capabilities/technology/data-analytics/">Visit our website.</a></h4><h4>Find us on <a href="https://www.facebook.com/BearingPointNorway/?ref=bookmarks">Facebook</a> and <a href="https://www.instagram.com/bearingpoint_norge/">Instagram</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e033e80118c6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bearingpoint-data-analytics-ai/a-tailored-data-mesh-for-remarkable-e033e80118c6">A tailored Data Mesh for reMarkable</a> was originally published in <a href="https://medium.com/bearingpoint-data-analytics-ai">BearingPoint Data, Analytics &amp; AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Four tips for predicting and preventing churn]]></title>
            <link>https://medium.com/bearingpoint-data-analytics-ai/four-tips-for-predicting-and-preventing-churn-9d6091c5a7fe?source=rss-23175f7f65bd------2</link>
            <guid isPermaLink="false">https://medium.com/p/9d6091c5a7fe</guid>
            <category><![CDATA[customer-experience]]></category>
            <category><![CDATA[churn-prediction]]></category>
            <category><![CDATA[customer-analytics]]></category>
            <category><![CDATA[advanced-analytics]]></category>
            <category><![CDATA[analytics]]></category>
            <dc:creator><![CDATA[BearingPoint Data, Analytics & AI]]></dc:creator>
            <pubDate>Mon, 28 Mar 2022 11:38:47 GMT</pubDate>
            <atom:updated>2022-06-22T12:10:35.317Z</atom:updated>
            <content:encoded><![CDATA[<p>By Jarrod Carmichael (Technology Analyst), Alborz Sabetrasekh (Technology Analyst), Vebjørn Axelsen (Partner) and Dr. Lars Bjålie (Technology Advisor) at BearingPoint</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jlHiBGs39Pc3Sq8ZjHEk7Q.png" /></figure><h4>Is predictive churn modeling a new thing? For sure not. The loss of profitable customers is an inherent challenge in all competitive industries, and for many companies this has been among the very first application areas for machine learning. Still, we see repeated examples of such initiatives not meeting expectations when it comes to actual improvements in customer retention.</h4><p>So, how to do better? We have ourselves been through significant learnings over the past five years, building on constant developments in business know-how, data availability and machine learning capabilities. In this article, we share some key steps to successfully implement a churn prediction model and actually reduce churn, based on our experience in industries such as banking, media and telecom, as well as a master thesis written by two of the authors in collaboration with a large Nordic financial institution.</p><p>Before diving into our churn prediction tips, we first encourage you to consider the churn problem holistically. An important starting point is to have control over target figures such as churn rates and customer profitability to quantify the problem and the financial upside in reducing the churn rate. Once that is in place, we recommend four categories of measures taking advantage of both analytics and domain expertise, as illustrated in the figure below. The topic of this article is highlighted in green; churn prediction, which is highly operational — taking actions to prevent churn on the individual customer level. Thinking more strategically, it is also possible to incorporate anti-churn measures in product development and marketing, customer journeys (which we have previously written an <a href="https://bearingpoint-analytics-ai.medium.com/how-behavioral-personas-can-transform-customer-journey-design-96035db3406">article</a> about), and finding churn drivers to re-engineer situations where churn occurs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*65BL31lrEZemdJ_GrrYJug.png" /></figure><p>Now, let’s get into our top tips for successful churn prediction.</p><h3>Clearly define churn</h3><p>One common pitfall when embarking on any analytics project is the lack of a clearly defined business problem. A problem definition sets the goal for the analytical model (i.e., defines the target variable we want our model to predict), describes how it will be operationalized, and gives structure to the business case. When working with customer churn this means that we need to clearly define what constitutes churn. Often, there may be multiple definitions of churn within the same line of business. We recommend that analytics and business resources work closely together — domain expertise is crucial to produce definitions that are both feasible to extract from the company’s data and have the potential for value creation, such as identifying profitable customers which can be influenced to stay.</p><p>There are several elements that should be considered in the definitions, most typically:</p><p><strong>1.</strong> <strong>The customer group we would like to keep</strong></p><p>We want to identify the customers we want to keep long-term. Ideally, we know each customer’s historic profitability, or even better, future estimated profitability (predicted customer lifetime value), and can filter out customers with negative profitability. If this information is not readily available, identifying heuristics for profitability such as tenure or total revenue might make sense.</p><p><strong>2.</strong> <strong>The churn event</strong></p><p>What constitutes churn? And what specific event do we want to predict? For a subscription-based service, churn is typically defined as a canceled subscription, but is it the customer <em>churn</em> <em>intent</em> that we want to predict, or is it the <em>actual churn</em> (i.e. actual deactivation of an agreement)? It depends on when we can counteract the churn, which determines the timing of the anti-churn measure. Within retail, can “no purchases within a certain timeframe” be the churn event? Ensure crystal clear definitions of these matters!</p><p><strong>3.</strong> <strong>The product connected to the customer and event</strong></p><p>This is often straightforward, but a complicated internal product structure can make this messy. Does it make sense to group together similar products (e.g., within banking, different long-term saving products), both to keep things simple and to have more historical churn events as examples to train our model on?</p><p><strong>4.</strong> <strong>The prediction window (i.e., time frame in which the event occurs)</strong></p><p>The most important consideration for deciding the length of the prediction window is how quickly the business can act on the prediction, which connects directly to the churn event and related anti-churn measure (point 2 above). A short time window will typically not allow sufficient time to contact the customer (and perhaps also be challenging in training a prediction model, as the fraction of customers in the training data set labeled “churn” decreases with the prediction window), whereas a long window will give a too general prediction (in the long term, all customers churn in some form or another!)</p><p><strong>5.</strong> <strong>Cause of the churn event</strong></p><p>Can we differentiate between customers who have churned willingly (which we can influence to stay) and customers who have stopped using the product or service due to some other external circumstance such as relocation or no longer having any need? Clearly, our anti-churn measures will be futile in the latter group.</p><p>We summarize these considerations in the below figure.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MUKAmFRgyy5dLfU8xCk-Ow.png" /></figure><h3>Use individual churn drivers to find the right actions</h3><p>After carefully defining the churn variable so that you can train, test and deploy an accurate prediction model, how should you reach out to the identified customers? What measures can you take to retain them? A special offer, a “check-in” conversation, or a lower price? An anti-churn project should include strong involvement of domain experts to shortlist possible actions, which are then piloted to test which work best (A/B testing of different offers for separate experiment groups, as explained in one of our <a href="https://www.bearingpoint.com/en-no/blog/predicting-behavior-and-taking-action-correctly-designing-ab-tests/">previous article</a>s).</p><p>It is possible to make this process more efficient by finding churn drivers for each individual customer, answering the question “<em>why will this customer churn?”</em>. This can help narrow down what actions are worth testing and create differentiated customer groups for even more accurate targeting of actions. Essentially, we would like to know how important each feature (data variable describing a customer characteristic, behavior etc.) is in making the specific prediction, and in what direction (lower/higher churn probability) the feature value contributed to the prediction. We want to know this independently of our choice of machine learning model, as less explainable “black box” models such as XGBoost and deep neural nets are likely going to give the best model. One way to achieve this is SHAP (Shapley Additive exPlanations), which we have recently used in several churn projects. We won’t go into the technical details here, but <a href="https://christophm.github.io/interpretable-ml-book/shapley.html">this</a> blog post provides an overview of the theory behind SHAP. In python, the SHAP-tools we need are implemented in the <a href="https://shap.readthedocs.io/en/latest/index.html">shap</a> package.</p><p>The figure below shows a fictitious example of a SHAP <em>Waterfall Chart</em> for a single customer calculated from a churn model in a telecom company. The green boxes indicate an increase in the probability of churn, while the blue boxes indicate a decrease in the probability of churn, compared to a baseline probability of 0,18.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*azKHWBeHVnKyzSj7ujKEIQ.png" /></figure><p>This plot tells us that the top churn driver for this customer is frequent contact with customer service. This variable could be an indication of a customer that has experienced some churn-triggering event which required lots of interaction with customer service to resolve (e.g., loss of cell phone service), or that poor customer service itself could be driving churn. Discussion with domain experts who understand the operation of the business and have a good understanding of customer behavior and churn events is crucial to make the right judgment calls in such situations — do we understand the driver well enough to create an appropriate anti-churn action? Essentially, what we are after is to separate which features have a <strong>causal</strong> relationship with churn, or if they are merely <strong>correlated</strong> — the causal features are the ones you want to base your churn actions on!</p><p>Continuing with our example, let’s say we know for sure that the “number of contacts to customer service” is indicative of an underlying problem which causes churn (and frequent calls to the customer service department), and not in itself a cause of churn. Coming up with measures to decrease the number of times a customer calls customer service is thus not a good strategy for reducing churn, even though there is a clear correlation between a customer’s frequent contacts with the company and their propensity to churn! We will not delve too deeply into the issue of correlation vs causation, but we encourage the interested reader to read <a href="https://towardsdatascience.com/be-careful-when-interpreting-predictive-models-in-search-of-causal-insights-e68626e664b6">this article</a>, which explores some common pitfalls and explains approaches to infer causal relationships.</p><p>We also see that surpassing the limit for internet usage each month according to the subscription plan is a strong churn driver, which is clearly a causal feature and much simpler to understand: The customer is not on an appropriate plan, and is running out of data. Sending a text message with a link to change the plan may be exactly what’s needed in this case.</p><p>Lastly, tailoring anti-churn measures to each unique customer is of course not realistic, and not our suggestion. However, based on aggregating insights from driver analysis on the individual level, some repeating patterns typically emerge where typical causes for churn behavior can be understood. This again can be exploited to identify anti-churn measures, preferably supported by A/B testing of the different measures.</p><h3>Find the model’s optimal intervention level</h3><p>Engagement with key stakeholders throughout the anti-churn project is of particular importance. The impact of the model on a company’s bottom line needs to be analyzed and communicated clearly. One straightforward way of doing this is by assigning an assumed income/cost for each customer in the model’s four different prediction categories:</p><p><strong>True positive (TP)</strong>: Customers correctly predicted to churn. For each of these customers, we have some probability of preventing churn, which will give us increased income from a longer customer lifetime. However, there is also a cost associated with the intervention.</p><p><strong>False positive (FP)</strong>: Customers incorrectly predicted to churn. The company incurs intervention costs without any churn reduction.</p><p><strong>True negative (TN)</strong>: Customers correctly predicted to not churn. No action taken, so no cost or income is associated with these customers.</p><p><strong>False negative (FN)</strong>: Customers incorrectly predicted not to churn, which we do not contact even though we ideally should. The company loses out on potential income from preventing churn, but we do not count this as a cost associated with operationalizing the churn model.</p><p>This is summarized in the cost matrix below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pvVhO0UtMY7DfmdMSAVbAA.png" /></figure><p>The assumed incomes/costs for the TP and FP categories together with the number of customers in each category will let us estimate the total profitability of operationalizing the model. The distribution of the customers in the four categories depends on the <em>intervention level</em> — the minimum model score required for classifying a customer as “Believes Churn” (TP+FP). This tells us which intervention level will give us the maximum profitability, by plotting profitability versus intervention level. If we have trained different models (i.e., tested different algorithms), the curve also tells us which model will give us the maximum profitability overall — an effective way to give the business side some insight into the model development process!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QAwfA-kLg25C2CfRcfEUqg.png" /></figure><p>This type of chart can be calculated for different sets of cost/income assumptions for the four categories, based on discussions with domain experts. Typical assumptions are the average cost of targeting a customer, the probability that a customer responds positively to a specific type of measure, and how long the customer remains after a successful measure. Typically, no clear answers to these questions are found, and undertaking a pilot project based on an A/B-testing framework before fully committing will reduce risks.</p><h3>Enrich your churn model with several sources of data</h3><p>A final best practice is to enrich your predictive churn model with several data sources. Time is virtually always better spent hunting down and structuring new data sources into features, than improving the model itself — the essence of the <a href="https://www.forbes.com/sites/gilpress/2021/06/16/andrew-ng-launches-a-campaign-for-data-centric-ai/">Data Centric AI manifesto</a>. In addition to potentially picking up more relevant signals statistically linked to customer churn, more features, up to a certain point, will also make your model more explainable.</p><p>The obvious place to start is of course the company’s own internal data, typically sourced from CRM or ERP systems, such as demographics, product data, and customer interactions. We would like to highlight some especially important features for churn modeling:</p><p><strong>Net promoter score:</strong> Perhaps one of the most indicative measures of customer satisfaction can be obtained by simply asking the customer about their perception of your company’s products and/or services. You might have received a text message after calling a customer call center, asking you to give a score on whether you are likely to recommend a product or service to a friend or colleague, followed up by an encouragement to leave a comment. This is commonly referred to as the “Net promoter score” and is a widely used metric in marketing research. An obvious benefit to using such a survey is that it allows your business to gain deep insights into the exact pain points of your customer journeys, which can be used to produce actionable measures to counteract overall customer churn. The net promoter scores gathered from this survey can then be used to track customer satisfaction following any anti-churn measures, using the A/B testing framework mentioned earlier.</p><p>Lastly, information gathered from this type of survey may be used as an input to your predictive churn model, which together with the other variables in your model may provide an effective predictor of customer churn.</p><p><strong>RFM variables</strong>: <a href="https://en.wikipedia.org/wiki/RFM_(market_research)">RFM</a> stands for recency, frequency, and monetary value. These are behavioral factors that are widely used in marketing analytics and often turn out to be significant predictors of churn. In short, these can be defined as follows:</p><p><em>Recency</em>:<strong> </strong>How long ago did a customer make a purchase/use a service?</p><p><em>Frequency</em><strong>:</strong> How often does the customer, on average, make a purchase/use a service?</p><p><em>Monetary value</em><strong>:</strong> How much does the customer spend on a given product/service?</p><p>We also recommend looking outside the walls of the company, including external data sources where possible. For example, in the master thesis mentioned in the introduction, we utilized external data from financial markets in predicting the mutual fund outflows for a large Nordic financial institution. We found (as perhaps expected) that stock market uncertainty, measured in the <a href="https://en.wikipedia.org/wiki/VIX">volatility index</a> (VIX), had a large contribution to both model performance and explainability, allowing us to gauge how much uncertainty in financial markets increases the likelihood of churn. In addition, customers whose mutual funds performed poorly were more likely to churn if the customer’s stated reason for saving is short-term savings, compared to someone saving for retirement.</p><p>In short, considering which features to include in your model requires a blend of experience, creativity, domain knowledge, and a healthy dose of trial and error. Iterating with domain experts such as customer advisors or business developers is crucial. Talking directly to existing and previous customers is also often forgotten and can be a valuable source of information.</p><h3>Summary</h3><p>Leveraging data through machine learning to get a clearer picture of which customers are going to churn for what reasons makes good economic sense, as it is often more expensive to gain a new customer than to retain an existing customer. However, it is not as simple as just training a prediction model, contacting the customers that the model predicts will churn and hoping for the best. There are many pitfalls, and the success of your machine learning anti-churn project starts at the very top, with a clear definition of the business problem. When operationalizing the model you should consider the underlying business case, and tailor the churn-preventing communication to why the customer is likely to churn in the first place, instead of a “one size fits all”-approach. To achieve a sufficiently accurate model it can also be well worth the time to look for new data sources, often outside the four walls of the company. Finally, we also highly recommend having an <a href="https://www.bearingpoint.com/en-no/blog/predicting-behavior-and-taking-action-correctly-designing-ab-tests/">A/B testing framework</a> in place.</p><h3>About the authors</h3><p><a href="https://www.linkedin.com/in/jarrodcarmichael/"><strong>Jarrod Carmichael</strong></a> is a Technology Analyst at BearingPoint Oslo, with experience in machine learning in safety-critical systems within the energy industry and customer retention within the finance industry. Before joining BearingPoint in 2021, Jarrod obtained a master’s degree in Financial Economics and Business Analytics from BI Norwegian Business School. Email address: jarrod.carmichael@bearingpoint.com.</p><p><a href="https://www.linkedin.com/in/alborz-sabetrasekh-3bb16b1a0/"><strong>Alborz Sabetrasekh</strong></a> is a Technology Analyst in the Data &amp; Analytics team of BearingPoint Oslo, with several years of working experience in the financial industry, including involvement in multiple projects related to customer retention. Alborz obtained a master’s degree in Business Analytics from BI Norwegian Business School in 2021. Email address: <a href="/service/mailto:alborz.sabetrasekh@bearingpoint.com">alborz.sabetrasekh@bearingpoint.com</a>.</p><p><a href="https://www.linkedin.com/in/vebjornaxelsen/"><strong>Vebjørn Axelsen</strong></a> is a Partner working at BearingPoint in Oslo and head of the Advanced Analytics area. He has significant experience as strategic advisor, architect and developer across data science and data engineering and has led a multitude of data science efforts across industries. Vebjørn obtained his MSc. in Computer Science specialized in AI from the Norwegian University of Science and Technology in 2007. Email address: vebjorn.axelsen@bearingpoint.com.</p><p><a href="https://www.linkedin.com/in/lars-bj%C3%A5lie-ba7b7754/"><strong>Dr. Lars Bjålie</strong></a> is a Technology Advisor at BearingPoint Oslo, with experience in customer analytics from the financial industry and predictive maintenance within railroad infrastructure. Before joining BearingPoint in 2016, Lars obtained a Ph.D. in computational materials science from the University of California, Santa Barbara.</p><h3>Want to find out more? <a href="https://www.bearingpoint.com/en-no/our-expertise/capabilities/technology/data-analytics/">Visit our website.</a></h3><h3>Find us on <a href="https://www.facebook.com/BearingPointNorway/?ref=bookmarks">Facebook</a> and <a href="https://www.instagram.com/bearingpoint_norge/">Instagram</a>.</h3><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9d6091c5a7fe" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bearingpoint-data-analytics-ai/four-tips-for-predicting-and-preventing-churn-9d6091c5a7fe">Four tips for predicting and preventing churn</a> was originally published in <a href="https://medium.com/bearingpoint-data-analytics-ai">BearingPoint Data, Analytics &amp; AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Predicting behavior and taking action: Correctly designing A/B tests]]></title>
            <link>https://medium.com/bearingpoint-data-analytics-ai/predicting-behavior-and-taking-action-correctly-designing-a-b-tests-d2454a670490?source=rss-23175f7f65bd------2</link>
            <guid isPermaLink="false">https://medium.com/p/d2454a670490</guid>
            <category><![CDATA[a-b-testing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[advanced-analytics]]></category>
            <dc:creator><![CDATA[BearingPoint Data, Analytics & AI]]></dc:creator>
            <pubDate>Tue, 12 Jan 2021 07:54:17 GMT</pubDate>
            <atom:updated>2022-06-22T12:12:27.387Z</atom:updated>
            <content:encoded><![CDATA[<p>By Karine Foss (Technology Consultant), Stian Mikelsen (Senior Manager) and Dr. Lars Bjålie (Technology Advisor) at BearingPoint in Norway</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rcr6ABE2nogugdmuFq61lQ.jpeg" /></figure><p>Machine learning models to predict customer behavior and A/B testing of marketing actions have become key capabilities in any modern marketing function. Together, they increase the impact of the marketing spend by finding the right messages to deliver to each customer to increase sales and loyalty. In our consulting practice, we unfortunately often see the techniques used together in ways that defeat the value of one or both.</p><p>In this article, we share some key learnings rooted in both theory and practice on properly combining them: How to select experiment groups, evaluate the results, and determine statistical significance. The principles we describe are general and apply equally well within for example predictive maintenance, but we keep our discussion within the realm of customer analytics.</p><h3>The separation of selection and action</h3><p>Whether one uses cutting-edge machine learning algorithms trained on vast quantities of customer data, or business rules built on well-aged gut feeling, the essence is to find customer groups who have a high likelihood of behaving in a certain way and influence them by sending the right message.</p><p>Typically, we want to monitor two things:</p><ol><li>How accurate is the analytical model, i.e. how good is the model at selecting a group of customers who are likely to behave in a certain way? In other words, are we able to predict behavior?</li><li>How effective is the action we take to influence this behavior — are we taking the optimal action for the selected group? In the domain of human behavior, it can be very hard to find the right action, and the best action now might not be the best action next month.</li></ol><p>For example, within banking one can build a model to select customers who have a high probability of changing their mortgage to another provider, and then tailor a customer journey with specific offers designed to increase loyalty within the selected group. When the model’s selection ability decreases, the model should be retrained based on fresh data, or we need an entirely new model. The same principle goes for action effectiveness, where an ineffective action needs to be redesigned. We, therefore, want to implement A/B-testing of both model and action, such that different models (champion/challenger) and different actions (classical A/B) can be evaluated over time. With the right experiment groups, we can measure these two things independently and at the same time.</p><p>The approach we present is straightforward to implement and has been used in projects at several large Norwegian companies. It is based on a structured setup for continuous evaluation by using three or four different experiment groups (depending on the case at hand).</p><h3>Design of experiment groups</h3><p>Consider a customer churn example. A machine learning model is developed to predict which customers are likely to cancel their subscription/membership (aka. churn) in the future. To prevent the cancellation, some action is to be taken. Based on our machine learning model, we divide the customer base into mutually exclusive experiment groups, as illustrated in the figure below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XqO4GpMZsQzbYBsedG9k_A.png" /></figure><p>The customers the model believes will churn are in the “believes churn” group, while the rest are in the “believes not churn” group. Here we need to manually decide a cut-off, for example selecting the top 5% of customers with the highest scores, or customers with a sufficiently high probability of churn as estimated by the model. These considerations are part of the business case of the anti-churn activity, which we do not go further into here.</p><p>We then apply an action to prevent churn to a percentage of the “believes churn” group (A), while the rest (B) are left out as a control group. In the example, customers that received an action have a 10% churn rate, whereas those that did not have a 20% churn rate — the action reduces churn by 50% in the customers picked out by the model.</p><p>To measure the performance of the machine learning model, we need to compare the churn rates between the customers in the “believes churn” and “believes not churn” control groups who do not receive any action — groups B and D. The churn rates for these groups are 20% and 5%, respectively, and our example model thus finds customers four times more likely to churn than the rest.</p><p>We can also measure the effectiveness of the action in the “believes not churn” group by sending the action to a percentage of these customers (C) and comparing the churn rate with the rest of the customers in the “believes not churn” group (D). Doing this increases the complexity of the experimental setup, but if one sees a large positive effect by doing this comparison, it is a hint that one may want to decrease the selectivity of the model and include more customers in group A. In the extreme case, one may want to do away with the model entirely and send the action to <em>all</em> customers. In the example, an increased churn rate for group C is illustrated, meaning the action has a <em>negative </em>effect in the “believes not churn” group as more customers churn — better only target those the model selects!</p><p>To summarize, the model selects a customer group with a four times higher churn rate than the rest of the customer base, and the action on the target group reduces the churn rate by 50%. It is tempting to stop our analysis here, concluding that the model’s ability to select customers is good and that the action is effective. However, the result must be evaluated more carefully: We must check if our findings are statistically significant, or just random noise. A common approach to do so is to calculate the P-value of the experiment.</p><h3>P-values to test actions for statistical significance</h3><p>P-values are a way of explaining and quantifying the findings of a test; it tells us how likely it is for the result to be caused by chance. There are countless articles <a href="https://www.statisticshowto.com/z-test/">explaining P-values</a> so we will not explain the concept here. Similarly, there are many <a href="https://www.socscistatistics.com/tests/ztest/default2.aspx">online P-value calculators</a> you can use to test the validity of your results or you can set up an Excel-spreadsheet.</p><p>It is by no means a perfect tool, and there is no such thing as a black and white significant or non-significant outcome (the statistically inclined reader may for instance <a href="https://www.nature.com/articles/d41586-019-00857-9">read this article for a nuanced take on P-values</a>). However, P-values still have their usefulness, and we recommend using them to get an idea if the action caused an effect or the result was due to randomness. A P-value of 0.05 is often used to determine significance and means we have less than a 5% chance of the result being caused by chance.</p><p>Below we note some key learnings for performing tests for statistical significance:</p><h4><strong>1. We need large enough test groups</strong></h4><p>Returning to our mortgage churn example, if the target group was only 20 customers, 10 receiving the action and 10 reserved in the control group, then one customer would have churned in the action group A, and two customers in the control group B. This is hardly telling you anything, as this could easily be by chance. For a more realistic test, say the two groups were 100 customers each, so that 10 customers churned in group A and 20 in group B. Then the test statistic is -1.9803 and the corresponding P-value (for a two-tailed test) is 0.0477 &lt; 0.05, and you would conclude that the result is statistically significant.</p><h4><strong>2. The ideal is a 50/50 split between action and control</strong></h4><p>Which customers end up in the target group is selected by our model, and within the target group, we decide on the number of customers receiving the action (A) or being selected for control (B). Very often one faces organizational pressure to operate with a smaller control group to reach more customers with the action, but it is crucial to stress the importance of statistical significance.<em> </em>It is also possible (and encouraged!) to send the action to the customers in the control group once the effectiveness of the action on the target group has been established.</p><h4><strong>3. Keep the experiments clean</strong></h4><p>Each customer should be subjected to only one test (unless the tests are undeniably independent) at the same time. The control group for one action must also be kept clear of similar actions. If the customer base is large, a tidy solution is to divide it into several non-overlapping groups. In the below figure, this setup issue is demonstrated for a case where three A/B tests with similar actions are run in parallel. For Test 1, a separate part of the customer base is used, so the results are not affected by the other tests. For Tests 2 and 3, however, the same customer base is used, resulting in a very small control group (customers who didn’t receive any action at all), and the effects of one test are hard to distinguish from the other.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/901/1*qkQSdgtDZLsy7YC5FzBvyg.png" /></figure><h4><strong>4.</strong> <strong>Keep the experiments disciplined</strong></h4><p>Before running the experiment, decide on the number of customers in each of the groups and which outcomes you will measure (i.e., clicks, e-mail opens, or purchases). Failing to do so may inadvertently lead to invalid results: A test giving a P-value below 0.05 means that there is a 5% chance that the action was not significant. If you “stack” multiple tests, it is easy for one of them to randomly give a P-value below 0.05 — this is known as <em>P-hacking</em>. Within customer analytics, it is tempting to run the experiment and expose customers to the action <em>until you reach significance (P &lt; 0.05), </em>which is a form of P-hacking. The probability of the observed significance being caused by random chance dramatically increases beyond 5%, and the result will be misleading (<a href="http://blog.analytics-toolkit.com/2017/the-bane-of-ab-testing-reaching-statistical-significance/">see this blog</a> for a more detailed explanation). There exists several techniques to overcome these pitfalls, but these are beyond the scope of this article.</p><h3>Practical considerations</h3><p>The approach we have described may seem straightforward to the statistician, analyst, or data scientist. However, in our consulting practice, we have encountered and overcome several practical challenges that may also be new to this group. Here we present our top tips for running experiments in real life:</p><h4><strong>1. Implement a scalable measurement architecture</strong></h4><p>We need to have a plan for collecting and storing data for our experiments. Important things to consider are how the customers in the experimental groups are labeled, and how the outcomes (product bought, link clicked, etc.) are registered. Also, if the organization wants to run many A/B tests in parallel, a system based on policies making sure customers are not subjected to several tests influencing each other (as discussed in the previous section) is also necessary. Building a common infrastructure to tackle these issues and automatically allocate customers to the different experimental groups and calculate effect and statistical significance is highly recommended. This ensures both scalability for running multiple tests, automating away manual work as much as possible, and makes sure that the integrity of each test is upheld by applying a common logic.</p><h4><strong>2. Clarify business goals and expectations</strong></h4><p>Oftentimes, the data science function meets resistance from the sales/marketing function. Below we list some typical situations we have experienced and questions that may be worth considering if they come up:</p><p><strong>a. Let’s skip the control group — this is important and we are sure our action has an impact</strong><em><br></em>Removing the control group prevents the measurement of the impact — wouldn’t it be nice to know just how much impact was generated, to argue for more marketing budget next year? And are you sure that the<em> </em>action couldn’t be optimized further?</p><p><strong>b. Wow! The action worked really well on the customers picked by the model. Let’s send it to everyone! </strong><em><br></em>Perhaps counterintuitive from a statistical standpoint, it makes perfect sense from a pragmatic point of view. The effort to create the action (i.e. copywriting, image selection, configuration in campaign management tool, etc.) has already been expended, and the business may not have any other relevant messages to send to their customers — something is often better than nothing. However, if the action involves some special offer, the business case may not be viable when sending to all customers, and not just those selected by the model.</p><p><strong>c. Even though we failed to determine that the action produced a statistically significant impact, or the impact was small, we should keep doing it</strong><em><br></em>Like the previous situation, it may still make sense to keep doing the action from a pragmatic viewpoint. However, there are also costs associated with continuing any action as it requires maintenance to stay up to date. For example, if the details of an advertised product change, the message has to change as well, and the model/business rule used to generate the target group may have to be updated if data sources change.</p><h4><strong>3. Seek long term business value and culture change instead of quick results</strong></h4><p>In the day-to-day work of designing and implementing tests, it is easy to lose focus on the long-term goal of measuring effect: increasing business value. Appreciate that a negative or inconclusive test gives useful information for the future; if an action is not effective, the business can focus their time and energy elsewhere. When communicating the results, have a plan for the alternative next step ready, and focus on what the results tell us instead of what they do not — take clear leadership in the development of an iterative, data-driven culture in the whole organization.</p><h3>Summary</h3><p>We have explained how predictions of behavior and the effect of marketing actions can be tested independently by correctly designing A/B tests. There are several practical challenges to be aware of, often arising from business wants and needs. In the desire for quick results, experimental setups may become faulty. Clarifying expectations, keeping a simple and clear experimental setup, and critically evaluating the results from a statistical standpoint will help develop better solutions giving long-term business value.</p><h3>About the authors</h3><p><a href="https://www.linkedin.com/in/karine-hages%C3%A6ther-foss-40511216b/"><strong>Karine Foss</strong></a> is a Technology Consultant at BearingPoint Oslo, specializing in Data &amp; Analytics, where she has experience with A/B testing within the media industry. She has a Master’s degree in Applied Physics and Mathematics, specializing in Statistics, from the Norwegian University of Science and Technology. Email address: <a href="/service/mailto:karine.foss@bearingpoint.com">karine.foss@bearingpoint.com</a>.</p><p><a href="https://www.linkedin.com/in/stian-mikelsen-2791621a/"><strong>Stian Mikelsen</strong></a> is a Senior Manager at BearingPoint Oslo, specializing in Data &amp; Analytics, where he has experience helping clients from big banks to start-ups build Data Platforms that provide real insight and value. He has a Master’s degree in Artificial Intelligence from the Norwegian University of Science and Technology. Email address: <a href="/service/mailto:stian.mikelsen@bearingpoint.com">stian.mikelsen@bearingpoint.com</a>.</p><p><a href="https://www.linkedin.com/in/lars-bj%C3%A5lie-ba7b7754/"><strong>Dr. Lars Bjålie</strong></a><strong> </strong>is a Technology Advisor at BearingPoint Oslo, with experience in customer analytics from the financial industry and predictive maintenance within railroad infrastructure. Before joining BearingPoint in 2016, Lars obtained a Ph.D. in computational materials science from the University of California, Santa Barbara.</p><h4>Want to find out more? <a href="https://www.bearingpoint.com/en-no/our-expertise/capabilities/technology/data-analytics/">Visit our website.</a></h4><h4>Follow us on <a href="https://www.facebook.com/BearingPointNorway/?ref=bookmarks">Facebook</a> and <a href="https://www.instagram.com/bearingpoint_norge/">Instagram</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d2454a670490" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bearingpoint-data-analytics-ai/predicting-behavior-and-taking-action-correctly-designing-a-b-tests-d2454a670490">Predicting behavior and taking action: Correctly designing A/B tests</a> was originally published in <a href="https://medium.com/bearingpoint-data-analytics-ai">BearingPoint Data, Analytics &amp; AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Six pitfalls to avoid when building a data-driven organization]]></title>
            <link>https://medium.com/bearingpoint-data-analytics-ai/six-pitfalls-to-avoid-when-building-a-data-driven-organization-1a26faabd2e1?source=rss-23175f7f65bd------2</link>
            <guid isPermaLink="false">https://medium.com/p/1a26faabd2e1</guid>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[data-driven]]></category>
            <category><![CDATA[advanced-analytics]]></category>
            <dc:creator><![CDATA[BearingPoint Data, Analytics & AI]]></dc:creator>
            <pubDate>Tue, 11 Aug 2020 10:29:32 GMT</pubDate>
            <atom:updated>2020-08-27T14:37:02.638Z</atom:updated>
            <content:encoded><![CDATA[<p>By Marcus Ehrndal (Head of the Data &amp; Analytics) at BearingPoint in Sweden</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5A5_AAzNHdsSlSJ6VvlcpQ.jpeg" /></figure><p><strong><em>More and more organizations are beginning to introduce a data driven approach into their ways of working. Despite their best intentions, a significant proportion will ultimately fail. Many are victims of the same pitfalls.</em></strong></p><p>In this post, Marcus Ehrndal, Head of the Data &amp; Analytics team at BearingPoint in Sweden, identifies some of the most common mistakes organizations make when attempting to adopt data and analytics into their ways of working and suggest ways you can avoid following in their footsteps.</p><p>There are numerous benefits to working in a data driven way. Here are our top three:</p><ol><li>Data driven decision-making is much more trustworthy than relying on your gut instinct</li><li>Using a holistic dataset rather than a smaller statistical sample to make decisions is more accurate and valuable</li><li>Automating elements of the decision-making process can drastically improve efficiency and be kinder on your bottom line</li></ol><p>At BearingPoint, we have an in-depth understanding of how to design, plan, and execute data and analytics initiatives. For more than 15 years we have partnered with some of the largest players in finance, insurance, and retail, as well as huge public infrastructure entities. These experiences have shaped our outlook on how to best utilize data and analytics and to identify when this resource is being used incorrectly.</p><p>Below is a list of some of the most common pitfalls to avoid when building a data-driven organization as well as some potential solutions. If you are doing any of the following, stop now, and get in touch with us!</p><p>1. Lack of strategic alignment — remember to focus on what value you need to create instead of what data you have available. At the end of the day, scaling requires buy-in from the rest of the organization. In order to make that happen, you need to ensure that you are using data to reach strategically important goals. These are the targets we recommend to prioritize.</p><p>2. Not defining how you intend to use the results and what value it will create — it is not enough to have a broad idea of the goal you aim to achieve. We recommend you to think beyond the initial start-up phase of the project and describe, in detail, how you will implement and operationalize the analytics model or dashboard you propose creating so that it actually produces discernible benefits. This sounds obvious, but it is a crucial step that is often overlooked until after a project has got well underway — by then it is often too late.</p><p>3. Technology overload — not picking the right tool for the job. While tempting, do not use advanced tools simply for the sake of it. In most cases, this will eventually limit or reduce the efficiency of the decision-making process that you are trying to facilitate. An elegant and simple solution is usually a lot simpler, cheaper, and easier for the average organization to use. Don’t overcomplicate things.</p><p>4. Waiting for the perfect infrastructure — don’t waste time waiting for the ideal opportunity to get started. Infrastructure and data quality, for example, can always be improved. But oftentimes, good, is good enough. Instead of waiting for the perfect moment, get your analytics project underway today.</p><p>5. Underwhelming UX — presentation matters. Your data and findings might be excellent but is the presentation of your results equally top-notch? If your dashboard or report looks terrible it can have a negative impact on what you are trying to communicate. Do not underestimate this point if you truly want the support, understanding and buy-in from the whole organization.</p><p>6. No traction — struggling with implementation. Your proof of concept might look like it’s going to be a success and bring value to the organization but once you try to make it an established process, you hit hurdle after hurdle. In the end, nothing tangible really happens. If you targeted a strategically important goal — as summarized in point 1 above — it will make it easier to implement the new processes because you should have backing from the organization. Fundamentally, no sustained value is gained until the data driven methodology is engrained in the process, so make sure to finish what you started instead of jumping from pilot to pilot.</p><p>As you can see, collecting and analyzing data without a clear goal in mind is risky behavior for every organization — and yet it happens. Perhaps in an attempt to get in on a good thing, many organizations get ahead of themselves before defining expectations, goals, uses etc.</p><p>The next time you initiate a data and analytics project, keep the above points in mind. Doing so will ensure that data collection and data analysis will be much more efficient, and the results will be more valuable.</p><p>If you are mid-project and some of the points sound a little too familiar, we encourage you to reach out before it’s too late. We are happy to discuss your current situation and suggest steps to get you and your organization back on track.</p><h3>About the author</h3><p><a href="https://www.linkedin.com/in/ehrndal/">Marcus Ehrndal</a> is Head of the Data &amp; Analytics team at BearingPoint in Sweden specializing in Data &amp; Analytics. His expertise is in the intersection between business and technology, and his career has been focused around the wide, complex, and ever more critical theme of creating business value from data. Email address: <a href="http://marcus.ehrndal@bearingpoint.com/">marcus.ehrndal@bearingpoint.com</a>.</p><h4>Want to find out more? <a href="https://www.bearingpoint.com/en-se/our-expertise/capabilities/technology/data-analytics/">Visit our website.</a></h4><h4>Find us on <a href="https://www.facebook.com/BearingPointSweden/">Facebook</a> and <a href="https://www.instagram.com/bearingpoint_sweden/">Instagram.</a></h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1a26faabd2e1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bearingpoint-data-analytics-ai/six-pitfalls-to-avoid-when-building-a-data-driven-organization-1a26faabd2e1">Six pitfalls to avoid when building a data-driven organization</a> was originally published in <a href="https://medium.com/bearingpoint-data-analytics-ai">BearingPoint Data, Analytics &amp; AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How behavioral personas can transform customer journey design]]></title>
            <link>https://medium.com/bearingpoint-data-analytics-ai/how-behavioral-personas-can-transform-customer-journey-design-96035db3406?source=rss-23175f7f65bd------2</link>
            <guid isPermaLink="false">https://medium.com/p/96035db3406</guid>
            <category><![CDATA[customer-journey]]></category>
            <category><![CDATA[advanced-analytics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[transformation]]></category>
            <category><![CDATA[customer-experience]]></category>
            <dc:creator><![CDATA[BearingPoint Data, Analytics & AI]]></dc:creator>
            <pubDate>Tue, 30 Jun 2020 08:59:27 GMT</pubDate>
            <atom:updated>2022-03-22T11:51:18.924Z</atom:updated>
            <content:encoded><![CDATA[<p>By Karan Kathuria (Senior Manager), Johannes Karlsen (Manager) and Vebjørn Axelsen (Partner) at BearingPoint in Norway</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W4qKf89Nwr1q3XDgvLJExA.jpeg" /></figure><h4><strong><em>When designing customer journeys to meet diverse customer expectations, what would you rather base your design on: Hypothetical guess-work about the types of customers you have, or actual insight from behavioral data?</em></strong></h4><p>There is little new about using personas to tailor marketing efforts and customer experience. However, traditional personas have relied on demographics and other personal descriptive data on customers in terms of “who they are”. This can be an effective way of humanizing the audience, but in today’s diverse environment these personas seldom capture the customers’ actual relationship with the product or service. Most would agree that a good customer journey is about tailoring the experience to the customers’ actual needs and expectations and not simply their age, gender or zip code. This is where behavioral personas come in handy, and in our view clearly outperform traditional approaches.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zJci_uBPegUm7bccFbtG8w.png" /><figcaption>Example segments (illustrative)</figcaption></figure><p>In 2019, BearingPoint collaborated with Plan Norge in redesigning the customer journey for their donors. The aim was to further improve their popular child sponsorship program (known as “Planfadder”) to drive increased loyalty and revenue.</p><p>In order to improve the customer journey, the idea of tailoring it to different types of donors quickly became the starting point. BearingPoint introduced the concept of behavioral personas which, in essence, are stereotypical descriptions of what donors actually do, as a basis to understand their needs and preferences. The underlying rational for this was that by understanding the different types of behavior we can see in interaction data between donors and Plan Norway and the sponsored child, we would be better able to understand the “what” and “why”, and thereby design a journey that would meet the donors expectations more effectively. This could be done by designing differentiated customer journeys, each well-tailored to observed patterns in actual customer behavior, and then map each donor to the best suited journey based on recent behavior.</p><p>The work was done by iteratively combining quantitative and qualitative methods. We used workshops to understand how todays customer journey could be improved and to identify relevant data to understand customer behavior. Machine learning algorithms on behavioral data were used to discover naturally occurring behavioral patterns in the customer base, which was a solid starting point for truly understanding the different types of donors. Interviews with actual donors enabled us to answer the “why” for each of the identified behavioral segments and enrich our understanding. In short, the results were highly convincing and can be summarized in two key take-aways:</p><h3><strong>1. Behavioral data enable a new level of understanding compared to demographic data</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AgrIMxWpUaESGh84JpRR4g.png" /><figcaption>Illustrative example</figcaption></figure><p>Behavioral personas, in combination with qualitative research, provide highly useful insights that answer the “what” and “why” rather than just “who”. These insights are valuable on a strategic level because they enable us to fundamentally improve and redesign a product or service. An illustration of this is how two of the validated groups of donors differed. Both where very concerned with giving money to charity, but where one group did this because of the relationship with the sponsored child, the other did not want to get emotionally involved with the sponsored child. This insight clearly demonstrated how important it is to understand the “why” if the goal is to improve the customer journey.</p><p>Interestingly, when analyzing how identified behavioral personas were distributed across different demographic variables such as age and postal codes, we found no clear patterns. The general conclusion was therefore that donor behavior had little to do with demographics, and consequently that donor behavior and expectations cannot be understood or predicted based on demographics.</p><h3><strong>2.</strong> <strong>Successful use of advanced analytics is all about applying the right techniques smartly</strong></h3><p>By neatly combing advanced analytics, domain expertise and qualitative research, we where able to gain substantial insight even by using “rather simple machine learning models”, as one of our AI-experts put it. The hardest part of the project was probably to decide which behavioral attributes to use for segmentation, which is all about understanding which differences in customer behavior are relevant for which types of customer journey differentiation, rather than just using a lot of data. By having a clear understanding of what we wanted to do with the insight, we were able to develop very specific and actionable donor segments within a few short iterations. When conducting interviews with donors we were surprised of how well the actual donors matched with their segment and how useful these multidimensional behavioral stereotypes were in sorting out the “why” among all of Plan’s donors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7vyVxLugHEXjxmdnUtX5Dw.png" /></figure><p>For the client, the project provided an eye-opener: A structured combination of a precise goal and problem definition, advanced analytics and qualitative research can be a both simple and very powerful approach to truly understanding your customers. Having this understanding in place really transformed the way Plan redesigned their customer journeys: Now the journeys are differentiated based on true insight into the needs and expectations of different types of customers, rather than on the often hypothetical personas we still see a lot of elsewhere.</p><h3>About the authors</h3><p><a href="https://www.linkedin.com/in/karan-kathuria-8ab38514/"><strong>Karan Kathuria</strong></a><strong> </strong>is a Senior Manager at BearingPoint in Oslo specializing in Data &amp; Analytics. His expertise lies in helping businesses make better decision by leveraging analytics and establishing a data driven way of working. He has hands-on experience from several industries and two AI startups. Karan obtained his MSc. in Industrial Economics and Technology Management from the Norwegian University of Science and Technology in 2011. Email address: karan.kathuria@bearingpoint.com.</p><p><a href="https://www.linkedin.com/in/johannes-karlsen-93283178/"><strong>Johannes Karlsen</strong></a><strong> </strong>is a Manager at BearingPoint in Oslo. Johannes is specialized in strategic use of Data &amp; Analytics and technology investments to improve customer value creation. Before joining BearingPoint, Johannes obtained his MSc. in Finance from the Norwegian School of Economics in 2017. Email address: johannes.karlsen@bearingpoint.com.</p><p><a href="https://www.linkedin.com/in/vebjornaxelsen/"><strong>Vebjørn Axelsen</strong></a> is a Partner working at BearingPoint in Oslo and head of the Advanced Analytics area. He has significant experience as strategic advisor, architect and developer across data science and data engineering and has led a multitude of data science efforts across industries. Vebjørn obtained his MSc. in Computer Science specialized in AI from the Norwegian University of Science and Technology in 2007. Email address: vebjorn.axelsen@bearingpoint.com.</p><h4><strong><em>Want to find out more? </em></strong><a href="https://www.bearingpoint.com/en-no/our-expertise/capabilities/technology/data-analytics/"><strong><em>Visit our website.</em></strong></a></h4><h4><strong><em>Find us on </em></strong><a href="https://www.facebook.com/BearingPointNorway/?ref=bookmarks"><strong><em>Facebook</em></strong></a><strong><em> and </em></strong><a href="https://www.instagram.com/bearingpoint_norge/"><strong><em>Instagram</em></strong></a><strong><em>.</em></strong></h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96035db3406" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bearingpoint-data-analytics-ai/how-behavioral-personas-can-transform-customer-journey-design-96035db3406">How behavioral personas can transform customer journey design</a> was originally published in <a href="https://medium.com/bearingpoint-data-analytics-ai">BearingPoint Data, Analytics &amp; AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Next Best Offer when you have few products but lots of data]]></title>
            <link>https://medium.com/swlh/next-best-offer-when-you-have-few-products-but-lots-of-data-521349035a9d?source=rss-23175f7f65bd------2</link>
            <guid isPermaLink="false">https://medium.com/p/521349035a9d</guid>
            <category><![CDATA[data]]></category>
            <category><![CDATA[crm]]></category>
            <category><![CDATA[advanced-analytics]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[BearingPoint Data, Analytics & AI]]></dc:creator>
            <pubDate>Mon, 15 Jun 2020 07:42:23 GMT</pubDate>
            <atom:updated>2022-06-22T12:13:21.188Z</atom:updated>
            <content:encoded><![CDATA[<p>By Tallak<strong> </strong>Bergheim<strong> </strong>(Manager), Vebjørn<strong> </strong>Axelsen (Partner) and Dr. Lars Bjålie (Technology Advisor) at BearingPoint in Norway</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JkFifK-Wgsd1Dzos8wOoRw.png" /></figure><blockquote>Which product is the most relevant to offer your customer right now?</blockquote><p>This is a central and repeating question within sales and marketing. Recommendation engines are often used to find the answer, and are key to the success of virtually all digital leaders such as <a href="https://hbr.org/2020/01/competing-in-the-age-of-ai">Amazon and Alibaba</a>. Typically building on identifying patterns in historical data about the sales of thousands of products, recommendation engines are known to perform well in retail settings. But what do you do if you don’t have that many products? And shouldn’t you be able to take advantage of customer data beyond sales, such as the customer’s dialogue with your business, product usage, payment information and demography?</p><p>In this article we describe an alternative approach for Next Best Offer (NBO), tailored to businesses with a much smaller number of products, such as banking, insurance and telecom. The core idea consists of mass produced and calibrated prediction models, calculating the probability of purchase for each individual product per customer. The method is based on projects delivered at several large Norwegian companies.</p><h3>Digital sales require knowing the Next Best Offer</h3><p>In the pre-digital era, sales personnel normally served a local group of customers tied to the geographic location of the store. This allowed for tailored product offerings based on an intimate understanding of the shop’s typical customer. For businesses such as banks and insurance companies, sales representatives may also have had a long-term relationship with the customer, enabling even more customized advice. This close relationship has been challenged as the digital era has brought centralization of sales personnel and “one size fits all” approaches. Despite this trend, customers still expect a high service level and to be recommended relevant offerings.</p><p>One important tool to resolve the tension between the customer’s high expectations and a centralized digital sales organization is recommendation engines. These analytical models estimate which product or service has the highest chance of being relevant for the customer — the Next Best Offer (NBO). Everyone who has shopped online or has used streaming content providers such as Netflix are familiar with product offerings at the end of the check-out, suggesting that “people who bought this also bought that” or “since you watched this movie you might also like that movie”, which come from such models.</p><p>In addition, there are often a limited number of interactions with a customer, and these interactions are usually when the customer is mentally in purchasing mode. In many businesses, rule-based recommendations are already in place to take advantage of these interactions — it can be as simple as offering credit cards to all customers who don’t already have one every time he/she logs into the online/mobile bank, or offering car insurance through a partner when selling a car. Moving beyond simple business rules to analytical NBO models can both increase the probability of using these interactions to up- or cross-sell, as well as reduce the chance that the customer feels that he or she is exposed to overly aggressive marketing with irrelevant products.</p><h3>Analytical models for few products, but lots of data</h3><p>The most frequently used analytical approach, so-called “collaborative filtering” methods, takes advantage of the combination of both large number of products and customers. They predict which products a customer will be the most interested in based on the preferences of other customers who bought/liked similar products, and/or the customer’s previous content preferences (e.g. product types). This type of model only requires data on which customers bought which products and are relatively simple to train and operationalize. They form the basis for the recommendation engines used by for example <a href="https://www.amazon.science/the-history-of-amazons-recommendation-algorithm">Amazon</a>. However, they require a very large amount of data on customer/product purchases to produce good results (think millions of customer interactions and products), and do not utilize contextual data about the customer (demographics, how the products are used, payment history, contact with the customer center etc.).</p><p>In contrast, many industries bear little resemblance to the mass-retail cases where collaborative filtering shines. Examples are banking, insurance and telecom, which usually have a product portfolio with somewhere between ten and hundred offerings, but rich datasets about their customers. These require a different analytical approach to predicting the next best offer. A more modern and technically demanding approach is “reinforcement learning”, which tests different offers and learn from the response (interaction/purchase), optimizing what is displayed individually to each customer. In contrast to collaborative filtering methods, this approach can quickly make predictions for new customers and products by learning on the fly, as well as taking advantage of contextual data about the customer. However, it is substantially more difficult to train and operationalize.</p><p>Instead of these two approaches, we suggest an intermediate method balancing prediction accuracy and complexity, which we call the “Prediction model factory”. Here we use supervised classification models and model score calibration techniques to predict the probability of purchase for each individual product per customer. The method is based on projects delivered at several Norwegian companies, taking advantage of already existing infrastructure for developing prediction models for single products.</p><h3>The prediction model factory</h3><p>The figure below illustrates the general supervised learning process, where machine learning algorithms predict the probability of a customer’s action by learning patterns in historical data consisting of explanatory variables and known outcomes. In a nutshell, with the “prediction model factory” approach, you train an individual prediction model for each of the products and services you want to recommend to your customers. The resulting model scores are then calibrated across products to give comparable probability scores in order to rank the products or services for each customer (for the technically inclined reader, see <a href="https://scikit-learn.org/stable/modules/calibration.html">here</a> for example). To get a first implementation up and running, each model uses the same classification algorithm (e.g. gradient boosted trees) trained on a standardized dataset with explanatory variables (such as customer demographics, interaction with products and customer service, product purchases etc.), differing only in the model target variable representing each different product. This allows for a simple mass production of individual models, hence the name “prediction model factory”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mg-RlQ2gSzYIRvf45wp-Pg.png" /></figure><p>One can later improve each individual model by creating special explanatory variables and fine-tuning parameters. In addition to product recommendations on an individual level, it is also natural to use the insight from these models to identify the target customer group for isolated marketing activities, such as sending an e-mail to customers with a high probability for purchasing car insurance or signing up for a credit card. If your business has a data science team, it is highly likely that some such models are already implemented and operationalized, which also provides a great starting point for building a model factory.</p><p>Once the models have been deployed and the prediction model factory is up and running, each customer will have a probability for purchasing/using each product. However, this probability alone is often not enough to make a sound business decision on the next best product/service to offer. There might be a need to weigh in other factors, such as customer or product profitability, or even strategic aspects, e.g. ambitions to increase market share for specific products. Furthermore, there might be other reasons why certain products should not be recommended to a specific customer at a certain time, for example if the customer has recently received an offer for a similar product. This process is shown schematically in the figure below and is a key part of the technical implementation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2NejBUBKUxB2Zssy2UIKPw.png" /></figure><h3>Five critical success factors for implementing an NBO system</h3><p>In general, an NBO system is not a stand-alone product that can be bought off the shelf, but a solution that should be tailored to specific business needs. This holds true for the prediction model factory as well, and below we outline the most important preparation steps for a successful implementation.</p><h4>Clear business case</h4><p>In addition to the investment of time and money, development of NBO systems also require new business processes and adjusted or new activities for sales and marketing personnel. To gain organizational buy-in, the NBO system should be supported by a business case which considers estimated sales increase versus cost of development and maintenance of models as well as potential increased sales channel costs. Creating realistic estimates for increased sales is the most challenging part of this process. However, if one has previously operationalized a prediction model, this might work as good proxy for estimates. In that case, one should consider both the models’ ability to predict customer behavior as well as sales personnel’s ability to use the insight actively in their communication. Further, if there are any existing next best offer solutions (e.g. based on simple rules) these are also excellent starting points. Should none of these be in place, it is also possible to build parts of the business case on experience with similar initiatives outside of your company.</p><h4>Data preparation and quality assessment</h4><p>The most important ingredient in any analytical model is the data it is based upon. Therefore, consider whether you have customer data of sufficient history and quality structured and easily accessible for analysis. What contextual data do you have available about your customers which may influence purchase behavior, e.g. demography, dialogue with your company, or product usage? Having efficient ways to source, integrate, quality assure and structure these data for analysis is essential for success in implementing any analytical model, including NBO. In a less data-mature organization, the key is to start small and not take every complexity into account from the beginning, but rather aim for a “good enough” data set as a starting point. From there, gradually implement good data practices to expand. Last but not least, whether you are legally allowed to use the relevant data for the modelling, and if it is ethical do to so, must of course also be considered.</p><h4>Understanding communication channels and processes</h4><p>Looking beyond the analytical model, one should review which customer interactions can be optimized with an NBO system and whether processes for customer interaction should change. Are new channels going to be used? For which message and in which situation? Do you need a different strategy for outbound (e.g. email and outgoing phone-calls) and inbound (e.g. website banner, offer when the customer calls the business) channels? For example, a customer service representative might receive a recommendation on the screen via a CRM system when a customer calls customer service, enabling a natural situation where one can suggest the product or service with the highest score. This might be a great way to utilize the NBO model, but it might also mean that new business processes must be developed, involving staff training and change management.</p><h4>Preparing for operationalization and system integrations</h4><p>Having established a business case, prepared accessible data, and considered which communication channels can take advantage of the NBO system, the next step is to draw up the technical architecture for operationalizing the model and integrating with consuming systems. Typical things to consider are:</p><ul><li>Do you have access to an environment where prediction models can be developed, tested, deployed and retrained, e.g. an application server or a cloud environment?</li><li>Which systems consume the prediction results (e.g. CRM system or a website)? How are these systems integrated with your analysis environment? Do you need a new system to deliver the prediction results to end users?</li></ul><p>You should also consider whether your company has staff with the required technical competency to implement the project. If you already have a data science team, how are they collaborating with sales and marketing to ensure that their analytical models create value in practice?</p><h4>Measuring model accuracy and activity impact</h4><p>Lastly, when the NBO model is up and running, we need to monitor both:</p><ul><li>That the prediction models are sufficiently accurate (how well do we predict what product each customer is most likely to buy next, i.e. how good is the model at selecting for interested customers).</li><li>That the activity itself has an impact on customer behavior (are we able to influence customers to make a purchase they otherwise wouldn’t have at that time, and do we have a greater influence on customers with a high score from the model?).</li></ul><p>Together these make up the increased sales from the NBO system, and it is important to measure these effects separately to make the correct strategic decisions and secure future buy-in for other analytics initiatives. In order to do so, you need to consider the infrastructure used to collect data from the systems consuming the prediction results, and where the logic for the effect measurements should be placed.</p><p>The principles for measuring model accuracy and activity impact are general for any prediction/action system, and in a future article we will go deeper into detail and discuss practical considerations and challenges in implementing this successfully.</p><h3>Summary</h3><p>As an increasing amount of sales will happen digitally, the importance of NBO systems will only increase in the years to come, and successful implementations of such systems are already the competitive advantage of some of the <a href="https://hbr.org/2020/01/competing-in-the-age-of-ai">world’s leading companies</a>. In this article we have outlined how businesses with lots of customer data but relatively few products can get started building their own NBO system, and if you think your own business can utilize our approach we encourage you to contact us.</p><h3>About the authors</h3><p><a href="https://www.linkedin.com/in/tallakbergheim/"><strong>Tallak Bergheim</strong></a><strong> </strong>is a Manager at BearingPoint Oslo specializing in Data &amp; Analytics. His expertise lies in utilizing Data &amp; Analytics for the financial and utilities industries. Before joining BearingPoint, Tallak obtained his MSc. in Industrial Economics and Technology Management from the Norwegian University of Science and Technology in 2016. Email address: tallak.bergheim@bearingpoint.com.</p><p><a href="https://www.linkedin.com/in/vebjornaxelsen/"><strong>Vebjørn Axelsen</strong></a> is a Partner working at BearingPoint Oslo and head of the Advanced Analytics area. He has significant experience as strategic advisor, architect and developer across data science and data engineering and has led a multitude of data science efforts across industries. Vebjørn obtained his MSc. in Computer Science specialized in AI from the Norwegian University of Science and Technology in 2007. Email address: vebjorn.axelsen@bearingpoint.com.</p><p><a href="https://www.linkedin.com/in/lars-bj%C3%A5lie-ba7b7754/"><strong>Dr. Lars Bjålie</strong></a> is a Technology Advisor at BearingPoint Oslo. His practical data science experience lies within customer analytics and predictive maintenance. Before joining BearingPoint in 2016, Lars obtained a Ph.D. in computational materials science from the University of California, Santa Barbara.</p><h4>Want to find out more? <a href="https://www.bearingpoint.com/en-no/our-expertise/capabilities/technology/data-analytics/">Visit our website.</a></h4><h4>Find us on <a href="https://www.facebook.com/BearingPointNorway/?ref=bookmarks">Facebook</a> and <a href="https://www.instagram.com/bearingpoint_norge/">Instagram</a>.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=521349035a9d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swlh/next-best-offer-when-you-have-few-products-but-lots-of-data-521349035a9d">Next Best Offer when you have few products but lots of data</a> was originally published in <a href="https://medium.com/swlh">The Startup</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>