<?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 Arnaud Dessein on Medium]]></title>
        <description><![CDATA[Stories by Arnaud Dessein on Medium]]></description>
        <link>https://medium.com/@arnaud.f.dessein?source=rss-2bdcdf8d43c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*3t2WlNvQy0hHHkhpUzEPjg.jpeg</url>
            <title>Stories by Arnaud Dessein on Medium</title>
            <link>https://medium.com/@arnaud.f.dessein?source=rss-2bdcdf8d43c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 20:17:46 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@arnaud.f.dessein/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 for Product Managers (Part 2/2)]]></title>
            <link>https://engineering.backmarket.com/data-for-product-managers-part-2-2-2b90af98b275?source=rss-2bdcdf8d43c------2</link>
            <guid isPermaLink="false">https://medium.com/p/2b90af98b275</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Arnaud Dessein]]></dc:creator>
            <pubDate>Mon, 30 May 2022 09:35:52 GMT</pubDate>
            <atom:updated>2022-05-31T19:50:59.017Z</atom:updated>
            <content:encoded><![CDATA[<h4>Modern fundamentals of data you need to thrive in a digital product management role</h4><blockquote>This article is the sequel to <a href="https://medium.com/@arnaud.f.dessein/fd2967333c00">Data for Product Managers (Part 1/2)</a>. We explained how product managers leverage data by discovering actionable insights and developing data products to create business value. We focus here on machine learning products and provide closing notes with tips and takeaways as well as references for further reading.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4BXKvYwcZPqc7BULfCyCtQ.jpeg" /></figure><h3>3 — Machine learning products</h3><h4>Two stages of integration</h4><p>Machine learning products integrate 2 stages of modeling and activation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hwUOjR7-0X74VKPqA8qtpg.png" /></figure><p><strong>Modeling. </strong>In the first stage, we model the available data with an algorithm to provide predictions or recommendations. For example, in e-commerce, it can be an algorithm to recommend articles that a user is likely to buy, and in the field of smart city, it can be an algorithm to predict the demand for bicycles in a bike-sharing station. We measure the performance of the model with appropriate machine learning indicators to assess if the predictions are good or the recommendations relevant. Common metrics include:</p><ul><li><em>Accuracy:</em> percentage of correct predictions among predictions in classification</li><li><em>Precision:</em> percentage of relevant instances retrieved among instances retrieved in classification and information retrieval</li><li><em>Recall:</em> percentage of relevant instances retrieved among relevant instances in classification and information retrieval</li><li><em>F-score:</em> combines the precision and recall as their harmonic mean (also called F-measure)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/212/1*scV7CriPs5BrTkwBq0wAiQ.png" /></figure><ul><li><em>Mean absolute error:</em> average error between the <em>N</em> predicted values <em>ŷᵢ</em> and actual values <em>yᵢ</em> in regression (also called mean absolute deviation)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/148/1*b5B_fcV61nzk8swhH7yOwA.png" /></figure><ul><li><em>Mean squared error:</em> average squared error between the <em>N</em> predicted values <em>ŷᵢ</em> and actual values <em>yᵢ</em> in regression (or the square root of it for the root mean squared error)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/164/1*wShHz4ZznrpTbGosA1BpwQ.png" /></figure><p><strong>Activation. </strong>In the second stage, we activate the model predictions or recommendations for users in a dedicated interface so that they can take action. For example, in e-commerce, we could display the recommended articles on the storefront website in a carousel in the middle of the homepage or below the product being viewed on a product detail page. In the field of smart city, we could display the predicted number of remaining bikes in the respective stations on a map via a mobile app. We would then measure the performance with product and business indicators such as the add-to-cart or conversion rate and the number of monthly active users. This overall performance depends both on the model performance and the activation relevance.</p><blockquote>The 2 stages are intrinsically linked.</blockquote><blockquote>On one hand, the user interface for activation depends on the modeling performance. In the context of autonomous vehicles, for instance, early algorithms enabled self-parking only, then proposed semi-automatic driving where manual interventions were still needed, and later evolved toward fully-automated driving. In e-commerce, recommending compatible products often requires a human-in-the-loop to prevent unacceptable errors (e.g., selling a mobile phone with a non-compatible cover).</blockquote><blockquote>On the other hand, activation data in real life can be used as a feedback loop to evaluate algorithms and improve the modeling performance. For example, we would check that the recommended articles on an e-commerce website convert and ingest this data over time to progressively optimize the recommendation engine.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nu_DrSKvJQcssg83Yd1kaQ.jpeg" /></figure><h4>Development specificities</h4><p>Developing machine learning products requires adaptation to several specificities:</p><p><strong>Uncertainty.</strong> Machine learning is an exploratory world.</p><ul><li><em>Model performance. </em>It’s hard to know in advance if and when a target modeling performance can be attained. Moreover, models might perform well on training data but fail to generalize on real-world data in production, a phenomenon called overfitting. Thus, product managers need to balance time-to-market and long-term value by arbitrating between fast iterations with incremental optimization versus large projects with more ambitious goals to radically increase performance.</li><li><em>Time management. </em>Timelines and tickets are also less predictable, so we need to be cautious on estimation so as to manage expectations — not to overpromise and underdeliver.</li><li><em>Error handling. </em>Because algorithms are imperfect, user research and common sense can help assess the acceptance criteria and determine which failure cases are expected versus which ones won’t be tolerated (e.g., different tolerance between article recommendations and self-driving cars).</li></ul><p><strong>Scalability.</strong> Production operations are different between software development (DevOps) and machine learning (MLOps).</p><ul><li><em>From conception to production. </em>In general, we train and test prototype models on a personal computer or a local server, and we then deploy the fine-tuned model in production. This latter step is complex because the model needs live access to data via dedicated APIs or pipelines (e.g., predict traffic incidents based on real-time weather). This might also require specific machines with appropriate CPU/GPU resources depending on the models.</li><li><em>Complexity and speed. </em>Complex algorithms will not only take more time to develop but also more time to run. Therefore, product managers need to balance the trade-off between model complexity and execution speed. Depending on the acceptable level of response latency for the system, we might prefer a simpler algorithm that executes faster (e.g., microseconds for high-frequency trading, milliseconds for article recommendations, and seconds for GPS routing calculation).</li><li><em>Model updates. </em>Another aspect to consider is the frequency of model retraining (e.g., daily, weekly, or monthly). Models generally improve with more or recent data, but there is a trade-off between model performance and the effort plus infrastructure required to train and update models.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W8Iwg-wftbeMhwRNc0noUQ.jpeg" /></figure><h4>Machine learning basics</h4><p>Machine learning techniques are classically divided into 3 main kinds of problems:</p><p><strong>Supervised learning.</strong> We have a dataset with labeled data, and we train a model with an algorithm to predict future labels when we’ll see unlabelled data. When some labels in the training dataset are missing, we call it semi-supervised learning.</p><ul><li><em>Regression and classification. </em>We call it regression if labels take continuous values (e.g., housing prices) and classification if labels are distinct categories (e.g., animal species in images).</li><li><em>Feature engineering and selection. </em>Prediction of the labels in the output of the model is based on variables or features that we input into the model. For example, we can feed past values in a model for time series prediction, and we can also add further features (e.g., hour, day of the week, month, weather, products bought, movies seen, or songs played) depending on what features are relevant for the domain and task. This step is called feature engineering. There are also algorithms for feature selection that automatically optimize the set of features given.</li><li><em>Models and algorithms. </em>There are many existing models with associated algorithms to solve supervised problems such as linear regression and lasso or ridge variants, decision trees and gradient boosting, random forests, nearest neighbors, support-vector machines, neural networks and deep learning, or genetic algorithms.</li></ul><p><strong>Unsupervised learning.</strong> We don’t have a training set with labels, but we find structure and hidden patterns within data. These can be used either to extract data insights or to feed supervised learning methods as input.</p><ul><li><em>Clustering. </em>This is called clustering when we form subsets of data points that show similarities within subsets but dissimilarities between subsets (e.g., create user segments or regroup similar products). Classical algorithms are k-means hard clustering and hierarchical agglomerative clustering.</li><li><em>Density estimation. </em>We also have density estimation where we fit a statistical model to explain data (e.g., estimate mean and standard deviation of normally-distributed points). Algorithms include maximum-likelihood estimation and expectation-maximization. They can also be used for soft clustering of data points, giving probabilities of belonging to the groups rather than strict assignment.</li><li><em>Dimensionality reduction. </em>Another problem is dimensionality reduction when we find redundancy in input data to summarize it with fewer variables (e.g., visualize high-dimensional data in 2D or 3D, decompose complex signals into simple atomic components). Known algorithms are principal component analysis, independent component analysis, non-negative matrix factorization, and multidimensional scaling.</li></ul><p><strong>Reinforcement learning. </strong>We have a computer program called agent that takes successive actions to navigate an environment so as to cumulate rewards. This is notably used in autonomous vehicles and in learning to play a game against a human opponent (e.g., go, chess, poker, or video games). The environment is often modeled as a Markov decision process that represents the succession of possible states and transition probabilities between states. Algorithms use dynamic programming techniques to explore these states and optimize rewards. Common techniques include the Monte Carlo method and Q-learning.</p><blockquote>The number one programming language for machine learning is Python with well-known compatible libraries for data processing (e.g., NumPy, SciPy, Matplotlib, Pandas, GeoPandas, and Airflow) and for machine learning in itself (e.g., Scikit-Learn, XGBoost, CatBoost, LightGBM, Prophet, PyTorch, Theano, TensorFlow, Keras, NLTK, and MLflow).</blockquote><blockquote>The machine learning community is large with many practitioners around the world and countless conferences, journals, books, and other resources. Notably, Kaggle is an online platform acquired by Google with more than 8 million registered users who can find and publish datasets, explore and build models, exchange and collaborate with peers, and enter competitions to solve data science problems with financial rewards.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qYff6lS8tGJgN5nh1vnOew.jpeg" /></figure><h3>4 — Closing notes</h3><h4>Eight shades of data jobs</h4><p>There is a wide spectrum of roles across the data world. In small companies, the frontier is often blurry because of limited resources and how everyone helps where needed. In larger companies, roles can be structured as follows:</p><p><strong>Data engineer</strong>. Creates extract-transform-load (ETL) pipelines to a data warehouse and reverse ETL pipelines to send data to third parties. Focused on the production-readiness of data.</p><p><strong>Data analyst.</strong> Performs data cleansing, data mining, and exploratory data analysis. Creates intelligible visualizations for product and/or business.</p><p><strong>Data scientist</strong>. Similar to a data analyst but with advanced skills in machine learning. Develops machine learning models for prediction and recommendation.</p><p><strong>Data product manager</strong>. Similar to a product manager but also focuses on data products. Discovers advanced data-driven insights.</p><p><strong>AI/ML engineer.</strong> Similar to a data engineer but also focuses on data pipelines for feature engineering. Deploys models in production.</p><p><strong>AI/ML scientist.</strong> Idem data scientist who focuses on machine learning.</p><p><strong>AI/ML researcher.</strong> Similar to an AI/ML scientist but also focuses on longer-term research and development. Sometimes writes articles and talks in scientific venues.</p><p><strong>AI/ML product manager.</strong> Similar to a data product manager but also focuses on machine learning products.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kmbzKGYXLhymCLN7vP7EJA.jpeg" /></figure><h4>Ethics of data</h4><p>The proliferation of data in digital products and, more generally, modern society has raised several ethical issues. Product managers have a duty to build responsible products by being conscious of and accounting for difficult questions:</p><p><strong>Bias.</strong> Algorithms, when trained on unbalanced datasets with few representatives of certain categories, can introduce biases in the outputs (e.g., gender discrimination and racism in facial recognition or hiring and lack of music diversity in song recommendations).</p><p><strong>Error.</strong> Algorithms can be wrong, which requires awareness to mitigate errors in critical situations and factual comparison to human mistakes (e.g., accidents involving self-driving cars can be more damaging and disputable than ones involving normal cars but are less likely to happen).</p><p><strong>Dilemma.</strong> Algorithms can face dilemmas when making decisions, and outputs sometimes depend on context and culture (e.g., deciding if a self-driving car should strike a child or an elderly person if a collision is inevitable).</p><p><strong>Interpretability.</strong> Complex models can be black boxes with no way to understand why they predicted or recommended outputs. This leads to questions on trust and the need for human interpretation (e.g., medical purposes when diagnosing a disease).</p><p><strong>Privacy.</strong> Personal data is tracked and monetized, which raises legal and security questions on data privacy. Laws like GDPR in the EU and CCPA in California have appeared as well as new actors for consent and preference management (e.g., OneTrust and Didomi) and controversies (e.g., Facebook–Cambridge Analytica data scandal and Google data breach).</p><p><strong>Environment.</strong> Data and algorithms consume massive resources that contribute to the global warming and ecological crisis (e.g., video and music streaming, data centers, cryptocurrency mining, and email and message volumes).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*upiE1X-ZssIJzYtNDk_dfw.jpeg" /></figure><h4>Tips and takeaways</h4><p>How can you apply all of this to your product role? One step at a time. Mastering data concepts for product management is a long journey. Try out a few elements incrementally and create your own path in this jungle.</p><p><strong>Develop a data-driven mindset. </strong>The first way to incorporate data and bring in business value is by discovering insights. Insights can come from market, user, or product research. They’re useful for making the right decisions and aligning people. Data also closes the loop to measure the impact of product developments with OKRs as a framework to structure strategic initiatives, KPIs and NSMs as metrics to measure impact, and the SMART method to define goals.</p><blockquote>So make sure to back up your ideas with research and insights rather than guesswork and intuitions. Set clear goals. Communicate your outcomes with relevant metrics. Often, people use metrics without really knowing what they mean. Data should not be a black box. Be the annoying person who asks questions and removes all ambiguity. Clarify the metric definition and single source of truth. Ask about the expected impact, how it will be measured, and the success criteria.</blockquote><blockquote>Talk with data teams to learn their language and issues. Learn standard SQL instructions, and understand CSV and JSON formats. Know the main data tables and tools used in your company. You don’t need to be a data product manager, but you should master the basics of analytics to be perceived as a legitimate expert in your scope and to challenge senior stakeholders. You can always get support from data teams on complex analysis.</blockquote><p><strong>Build machine learning products. </strong>The second way of leveraging data to bring business value is by developing data products. Data can be quantitative or qualitative. Applications increase in value and complexity, from descriptive and diagnostic to predictive and prescriptive. The capacity for machine learning products to create value depends on the performance of the 2 stages of modeling and activation. Such products have specificities in terms of uncertainty and scalability.</p><blockquote>Talk with data scientists to understand their world. Learn a bit of Python. Know how classical machine learning models, algorithms, and metrics work. Help determinate the relevant variables for feature engineering. You don’t need to be an AI/ML researcher, but you should develop a general understanding of machine learning.</blockquote><blockquote>Many companies market their product with buzzwords like big data and artificial intelligence only to find they don’t leverage those at all. There are also startups that focus on the feasibility of solutions for years with no user desirability or business viability to reach product-market fit and scale. Find the sweet spot between problems and solutions. Artificial intelligence can bring about new ideas, but be sure they address actual needs. Don’t over-engineer. A first predictive model for a minimum viable product can be as simple as a moving average or linear regression. Always focus on value like with standard products.</blockquote><p><strong>Think beyond data. </strong>Being data-driven is nothing without being user-centric. Best product managers enrich their storytelling by combining data insights and user empathy. Also, product managers should not be dogmatic about data. Some initiatives cannot be easily measured or supported by data when information is missing or takes too long to collect. We now talk of being data-informed rather than data-driven to convey this need for balance between data theory and practical reality. Sometimes, you don’t even need data when an idea is a no-brainer in relation to your mission, vision, and brand positioning.</p><blockquote>Take a leap of faith. Leaders are capable of making decisions amid uncertainty. Not all decisions are critical to the journey of your company. Focus on the few important decisions, collect as many insights as you can, and rely on your gut feeling for the unknown. For other decisions, remember that the status quo and inaction can cost more than making a suboptimal decision.</blockquote><blockquote>Don’t be afraid to move fast, be bold, fail, and learn. Don’t be paralyzed by risk. Don’t become that irrational rationalist!</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p6GXhDIYwQsvzxwbxUALzQ.jpeg" /></figure><h3>References</h3><ul><li><a href="https://medium.com/productschool/why-data-analytics-matters-for-product-managers-495149bb6936">Why Data Analytics Matters for Product Managers</a></li><li><a href="https://productcoalition.com/faq-what-skills-should-a-data-driven-product-manager-have-188b002cb5a5">FAQ: What Skills Should a Data-Driven Product Manager Have?</a></li><li><a href="https://itamargilad.com/the-tool-that-will-help-you-choose-better-product-ideas/">Idea Prioritization with ICE and the Confidence Meter</a></li><li><a href="https://medium.com/@CohortPlus/data-science-fundamentals-for-product-managers-4c347c67b337">Data Science Fundamentals for Product Managers</a></li><li><a href="https://blog.insightdatascience.com/an-introduction-to-the-data-product-management-landscape-ef930afe6de5">An Introduction to the Data Product Management Landscape</a></li><li><a href="https://medium.com/swlh/the-age-of-data-product-managers-how-to-prepare-24c0fedc163f">The Age of Data Product Managers: How to Prepare!</a></li><li><a href="https://towardsdatascience.com/data-product-management-ffa582f7e047">Data Product Management</a></li></ul><p>A big thanks to Alice Larue, Elizabeth Hudson, and Jean-Charles Sorin for reviewing this article!</p><p>If you want to join our Product team or any other Back Market department, <a href="https://jobs.lever.co/backmarket">take a look here, we’re hiring!</a> 🦄</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/741/0*uaYugO9kVlqF97K3.gif" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2b90af98b275" width="1" height="1" alt=""><hr><p><a href="https://engineering.backmarket.com/data-for-product-managers-part-2-2-2b90af98b275">Data for Product Managers (Part 2/2)</a> was originally published in <a href="https://engineering.backmarket.com/">Back Market Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data for Product Managers (Part 1/2)]]></title>
            <link>https://engineering.backmarket.com/data-for-product-managers-part-1-2-fd2967333c00?source=rss-2bdcdf8d43c------2</link>
            <guid isPermaLink="false">https://medium.com/p/fd2967333c00</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Arnaud Dessein]]></dc:creator>
            <pubDate>Thu, 28 Apr 2022 12:16:15 GMT</pubDate>
            <atom:updated>2022-05-30T20:34:24.480Z</atom:updated>
            <content:encoded><![CDATA[<h4>Modern fundamentals of data you need to thrive in a digital product management role</h4><blockquote>You’ve probably heard the word data a lot. How does it translate into your day-to-day as a product manager? Read this article to learn more and improve your product skills.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ibuei2nnPyxL10Plf_qwTw.jpeg" /></figure><p>Today, being user-centric isn’t enough for product managers. In modern digital product management, a data-driven mindset is required to make the right decisions and align stakeholders. Artificial intelligence is also ubiquitous in digital products of our daily life. All top companies now make extensive use of available data and follow various metrics to monitor performance, improve products, and increase business value. This article outlines the background on data to improve your skills and navigate this context with ease.</p><blockquote>In digital product management, data can be leveraged in 2 ways to create business value: discover actionable insights and develop data products.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4L8bLwnyPwMUKgBP8Y3l4Q.png" /></figure><h3>1 — Discover actionable insights</h3><h4>Three kinds of insights</h4><p>Insights can be divided into 3 main categories:</p><p><strong>Market research. </strong>Data that describes the market and uncovers demographic, economic, and statistical information about a specific industry. This can include market needs and size, competition analysis, brand positioning, consumer behavior and segmentation, and price sensitivity analysis.</p><p><strong>User research. </strong>Data that studies a small subset of people to understand who they are, what they want, what they think of a product, and where it can be improved. This can include user interviews, focus groups, card sorting, usability testing, and user feedback (e.g., likes, ratings and reviews, and net promoter scores).</p><p><strong>Product research.</strong> Data that understands how users interact with the product, where the product performs well, or where it should be improved. This can include tracking of web or mobile applications (e.g., views, bounces, clicks, and scrolls), user flows and funnels, heatmaps, and A/B testing. Tracking consists in code snippets that send events with properties (e.g., product added to the cart with ID, price, and brand) based on triggers (e.g., when the user clicks on the add-to-cart button). Events can be sent to external tools for analytics (e.g., Google Analytics, Adobe Analytics, Amplitude, Mixpanel, Contentsquare, or Hotjar) or to a database (e.g., Snowflake, Amazon Redshift, or BigQuery) for storing and later analysis via integrated functions or dedicated tools (e.g., Excel, Power BI, Tableau, Data Studio, Metabase, QlikView, or Airtable). Events can also be enriched with data from a data warehouse (DWH) and are typically centralized in a data management platform (DMP) or customer data platform (CDP) and activated in marketing automation tools (e.g., marketing campaigns, audience segmentation, personalization, or cross-channel orchestration).</p><blockquote>These 3 kinds of insights are useful and complementary in building a product and providing a complete story. Product insights become powerful when product usage is sufficient to reach statistical significance. Therefore, product insights might be harder for early-stage startups to obtain before product-market fit or for B2B products where the number of customers is usually smaller than scaled B2C products. In these situations, product managers need to rely even more on market and user research to perform relevant discovery and de-risk hypotheses.</blockquote><h4>Actioning these insights</h4><p>Data-driven product managers use the 3 types of insights to decide and align.</p><p><strong>Decide. </strong>Assess the opportunities and bets to address next. The goal here is to build the right product by solving the right problems and finding the right solutions. The basic prioritization method compares the estimated business or user impact (I) of an idea versus the technical effort (E) required to develop it. All insights are not equivalent. Product research data from the launch of a minimum viable product or an A/B test with millions of users is more reliable than user research with a couple of users or early-stage market insights. Hence, confidence (C) can also be used to mitigate uncertainties. Often, the impact is further computed as the reach (R) of users targeted (e.g., traffic per device, country, or page type) multiplied by the actual impact on these users. This gives the standard RICE score for feature prioritization and its variants. A confidence meter can also help quantify the confidence score depending on the insight categories.</p><p><strong>Align. </strong>Evangelize stakeholders and have them adhere to the product vision, strategy, and roadmap. It’s easier to convince people with factual, objective statements that are shared compared to personal, subjective ideas that can be disputed. Data-driven product managers move faster and remove individual blockers by engaging all stakeholders from business, design, tech, and executive teams. They’re good storytellers who can present compelling narratives that unite everybody. And they show the current realities and what could be improved, so people can dream of a better future. In a nutshell, data helps product leaders become visionaries who inspire!</p><blockquote>In the day-to-day of a product manager, bringing in data is powerful on 2 occasions: when pitching a new idea and when closing the loop of the build-measure-learn cycle of product iterations, where data illuminates the next steps — to either pivot (completely change the idea) or persevere (try out modifications).</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ECC2EZK-hiMw6DTcfGrQgg.png" /></figure><h4>Measuring the impact</h4><p>This is what closes the loop of product iterations and demonstrates the outcomes of a new product or feature. It’s important to think of the metrics and success criteria ahead. Here’s a lexicon of common acronyms to help set up goals in this context:</p><p><strong>OKR.</strong> Objectives and key results. Many tech companies have adopted and adapted the OKR framework popularized by Intel and Google. It consists of stating the main company objectives (e.g., launching a new product or successfully moving into a new market in 2022) and outlining measurable key results, typically 3 to 5, to achieve the respective objectives (e.g., reach 1,000 paid customers, keep customer acquisition cost below $10, and win 1 best-product-of-the-year award).</p><p><strong>KPI.</strong> Key performance indicator. They’re essential metrics for business and product (e.g., gross margin volume, customer lifetime value, return on ad spend, customer acquisition cost, number of app downloads, and monthly active users). We use them in the key results of the OKRs. It’s important to know their baseline values, compare them to market standards when possible, and follow them over time to monitor performance and detect product issues or market trends that require adjustment.</p><p><strong>NSM. </strong>North star metric. This is the central KPI of a product. It acts as a compass to ensure the product is heading in the right direction. It expresses the value for customers and represents the product vision and strategy. It’s actionable, measurable, and understandable so we can influence it, track it over time, and inspire the whole company (e.g., number of nights booked for Airbnb or daily active users for Facebook).</p><p><strong>SMART.</strong> Specific, measurable, achievable, relevant, time-bound. This mnemonic helps to set goals that specify an area for improvement. These goals should be measured, ambitious but reasonable enough to be achieved, relevant to the business, and bound in time (e.g., increase the net promoter score to 60 in Q3 of 2022). We also use them in OKRs. They’re useful for setting unambiguous goals for business, product, and people management.</p><h3>2 — Develop data products</h3><h4>Data science overview</h4><p>Data science emerged from the rise of artificial intelligence and its branch of machine learning. It has leveraged countless innovations over the last decades.</p><p><strong>Context. </strong>A field that uses computer science and mathematical methods to extract knowledge and apply actionable insights from data. Data science is pervasive within a broad range of applications such as recommendation (e.g., products, videos, music, contents, jobs, relations, and food), optimization (e.g., supply chain, stocks, and logistics), forecasting (e.g., travel time, weather, and crimes), recognition (e.g., images, faces, and songs), predictive maintenance (e.g., production lines, vehicles, and lifts), fraud detection (e.g., spams, cyberattacks, and insurance), and in domains including finance, marketing, retail, insurance, policing, healthcare, telecommunications, propriety, travel, mobility, sport, and entertainment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-bNEWW2jmKsgm7hasZT1zg.png" /></figure><p><strong>Technologies. </strong>Machine learning is not always perceived but has democratized and interfered with our lives, as shown in the non-exhaustive list of technologies below:</p><ul><li><em>Google:</em> page ranking for returning relevant links to search queries</li><li><em>Apple:</em> natural language understanding for Siri voice assistant</li><li><em>Facebook:</em> content and ad campaign optimization for personal feeds</li><li><em>Amazon:</em> natural language understanding for Alexa voice assistant</li><li><em>Microsoft:</em> natural language understanding for Cortana voice assistant</li><li><em>Tesla:</em> self-driving cars for transportation</li><li><em>Uber:</em> pricing and routing optimization for car rides</li><li><em>Airbnb:</em> pricing optimization and recommendation system for housing</li><li><em>Instagram: </em>content optimization and recommendation for personal feeds</li><li><em>LinkedIn:</em> content optimization and recommendation for personal feeds</li><li><em>Netflix:</em> recommendation for movies, series, shows, and documentaries</li><li><em>YouTube:</em> recommendation system for videos</li><li><em>Spotify:</em> recommendation system for music and podcasts</li><li><em>Shazam:</em> recognition system for songs</li><li><em>DeepL:</em> translation system between multiple languages</li></ul><h4>Two types of data</h4><p>In data science, we distinguish between 2 types of data:</p><p><strong>Quantitative data</strong>. These are numeric values, mostly including real numbers but also sometimes complex ones (e.g., spectrum for frequency analysis or magnetic resonance imaging). Real numbers can either be discrete, like integers (e.g., number of followers), or continuous (e.g., product price). They have a range (e.g., the age is a non-negative integer and the acceleration is a real number that can take positive and negative values). Discrete numbers with a finite range are called categorical, in particular, binary data when there are only 2 possible values (e.g., boolean encoded as 0/1 for false/true). Quantitative data has a given number of variables or dimension (e.g., GPS data shows two coordinates with respective ranges -90°/+90° and -180°/+180° for latitude and longitude, and RGB images are encoded with 2 spatial dimensions x/y plus 3 color dimensions red/green/blue per pixel). Data within a single dimension is called univariate, and multidimensional data is called multivariate. Common statistics for univariate data include the mean, median, standard deviation, min, and max.</p><p><strong>Qualitative data.</strong> These are non-numeric data, often taking the form of text or being transcribable into text (e.g., voice recording). It can be user-generated content (e.g., ratings and reviews for products and sellers, app store reviews, social network threads, and user surveys), recorded conversations with customers, and customer service comments. Generally, this is harder to process automatically via algorithms than quantitative data and is thus often analyzed manually in complement. There are, however, machine learning techniques that can encode certain textual data into quantitative data (e.g., surveys with a finite set of options) or automate natural language processing and natural language understanding tasks for more complex texts.</p><h4>Four levels of applications</h4><p>We have different levels of data science applications that might require the use of machine learning techniques.</p><p><strong>Levels.</strong> There are 4 levels of increasing value versus difficulty:</p><ul><li>Descriptive: factual description of what happened</li><li>Diagnostic: analysis of why something happened</li><li>Predictive: factual description of what will happen</li><li>Prescriptive: analysis of how to make something happen</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u9lCWbaxCtvryheeeVfMJA.png" /></figure><p><strong>With or without machine learning. </strong>For factual hindsight at the descriptive level or for manual insight at the diagnostic level, information is often represented in data products in the form of a dashboard with historical analytics computed on past data and filters on properties (e.g., dates, countries, device types, and user segments). Going further for automatic insight and foresight at the diagnostic, predictive, and prescriptive levels, machine learning is used to provide predictions and recommendations. Such machine learning products have important specificities to know.</p><blockquote>We detail the specificities of machine learning products and provide closing notes with tips and takeaways as well as references for further reading in the second part of this article <a href="https://engineering.backmarket.com/data-for-product-managers-part-2-2-2b90af98b275">Data for Product Managers (Part 2/2)</a>.</blockquote><p>A big thanks to Alice Larue and Elizabeth Hudson for reviewing this article!</p><p>If you want to join our Product team or any other Back Market department, <a href="https://jobs.lever.co/backmarket">take a look here, we’re hiring!</a> 🦄</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/741/0*uaYugO9kVlqF97K3.gif" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fd2967333c00" width="1" height="1" alt=""><hr><p><a href="https://engineering.backmarket.com/data-for-product-managers-part-1-2-fd2967333c00">Data for Product Managers (Part 1/2)</a> was originally published in <a href="https://engineering.backmarket.com/">Back Market Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>