Robyn Fadden – host: | Over the past several years, Montreal has become a hotbed of Artificial Intelligence innovation, in both business arenas and academic thought. Just recently, the International Research Laboratory was launched in Montreal to bring together organizations dedicated to artificial intelligence research in Montreal, including McGill University. That illustrates one of the reasons AI is thriving in Montreal: because of how valuable fundamental research and thought leadership is for understanding how AI’s numerous applications play out in the wider world, from analyzing healthcare data to implementing prescriptive AI as the future of data-driven decision making. Today on the Delve podcast, we’re narrowing that AI lens to retail, where inventory and customer data has grown massively in the online space. |
Robyn Fadden – host: | Welcome to the Delve podcast. I’m your host for this episode, Robyn Fadden. On this episode I talk with Desautels Professor Maxime Cohen, about demystifying predictive data analytics, so retailers can use it for operational decision making and to boost revenue. In his recent book, Demand Prediction in Retail – A Practical Guide to Leverage Data and Predictive Analytics, published by Springer this year, he and his co-authors (Paul-Emile Gras, Arthur Pentecoste, Renyu Zhang) provide a detailed guide for data scientists and students of business analytics. But the book also offers business insights to a wider audience. The information in the book points to something essential for all retailers today: how to leverage their historical data to predict future demand for their products. Welcome to the Delve podcast, Professor Cohen. Thank you for talking with me. |
Maxime Cohen: | Hi Robyn, thank you very much for having me today. We really hope that this book will help many retailers to do a better job at predicting demand and leveraging data in order to improve demand prediction |
Robyn Fadden – host: | To set the stage for our conversation, could you outline the current state of data analytics use in retail operations management? |
Maxime Cohen: | Definitely. Leveraging data analytics in the retail world is becoming more and more common and trendy. And what is interesting, it’s common for many types of different retailers, for example, from the smaller retailer all the way to the large corporation, it can be across different verticals like think about fashion, think about electronics, think about supermarkets, convenience stores. And also, it’s been applied both for online retail as well as for brick-and-mortar stores. Most of those retailers started collecting and storing some data. And the goal of analytics is to try to leverage this data and try to inform decisions, and to use the richness in that data to improve operational decisions for the business. |
Robyn Fadden – host: | And that data could come from customers, it could come from warehouses, it could come from all aspects of the organization. |
Maxime Cohen: | Exactly. This goal is to capture data from every single key stakeholder of the company. So of course, the customers, but it can also come from suppliers, from distributors, and so on and so forth. You really want to get kind of a 360 view of your operations through a data lens. |
Robyn Fadden – host: | Where does using predictive analytics fit into the grand scheme of how data analytics is applied for business purposes today? |
Maxime Cohen: | What I’d like to say about this topic is in data analytics, and AI, which stands for artificial intelligence, there are three levels of maturity. The first level is called descriptive analytics or descriptive AI. And there the goal is to answer the question of what happened in the past, you want to use the data, try to understand what happened, try to understand key interesting patterns and trends and identify customer segments, for example. The second level of maturity is predictive analytics or predictive AI. And there you try to answer the question of what will happen in the future. So you want to use past data to predict the future. And the biggest use case in that one is demand prediction in retail, which is the topic of the book. Well, you want to use the data from the past to predict the future demand of specific products at specific prices. And the third level of maturity is prescriptive analytics or prescriptive AI. And here the goal is to answer the question of how can I make optimal decisions in the future? So now we’re talking about prescriptions. We’re talking about recommendations, taking some actions, trying to answer the question how to decide key important decisions for my own business. Of course, this is a golden standard, you want to reach a certain level of maturity but different retailers have reached different levels. |
Robyn Fadden – host: | Why should retail businesses be using predictive data analysis at this point? And how common is it? |
Maxime Cohen: | It’s very important to nail down the second level of maturity, the predictive analytics before moving to prescriptive, because any prescriptive will need first to have good predictive foundations. And that’s why we spend a lot of time on the book to spend, and to try to see how retailers can leverage data to predict the demand. It’s going to highly depend on the company on the country on the size of the company and the appetite of the company to become more data driven and technology driven. Now, of course, smaller retailers have limited infrastructure and limited investment investments that they can put in that space, and therefore they’ll be a little bit more limited. And larger companies have already very good system very large infrastructure where they already are doing a lot of predictive analytics. What is interesting in the last few years, if you’re asking me the same question five years ago, I will tell you, it’s not that very common for retailers to use predictive analytics, but because of the pandemic is somewhat gave a wakeup call for retailers to start collecting more data and try to find ways to monetize this data to leverage this data to improve operational decisions in order to make more profits. |
Robyn Fadden – host: | Do you think that the pandemic made more data available as more people shopped online, providing more personal data than they would have when brick-and-mortar shopping? |
Maxime Cohen: | Exactly. So because more people move to the online channel, it’s much easier to collect data from online browsers. You can collect easy user data, as well as richer data, for example, you can have clickstream data, you know what people click on, so what they’re interested in. You can also know how long they spend in each web page. You can also know if they read some reviews, you can also know like, which part of the search on web search on the website, so you know what they are interested on. So that’s definitely one way. Another way is because there was a lot of retailers starting opening new online stores. So with this online store, it provides an opportunity to collect data, and therefore like a lot of retailers have recently moved to collecting more data. |
Robyn Fadden – host: | And now they need to figure out what they’re going to do with that data. |
Maxime Cohen: | That’s the next step. Like once you have the data, you store it, you collect it. And the next question is, how can you use it in a smart way to improve decisions? And in particular, in the book, we look at how retailers can use the data to predict demand. So you can ask, why is it important for predicting demand? Predicting demand is very important because it can guide many tactical and strategic decisions. For example, inventory management, you want to manage your inventory and your replenishment strategy, how much to purchase from each of the products on the shelves, how much to have in the warehouse, and when you will demand prediction, and to accurate level, you can make better inventory decisions. Another big decision is pricing and promotions, pricing and promotions, also depend on how well you are able to predict the demands. So predicting the demand can really help guide many decisions, inventory, supply chain, pricing and promotion, procurements, fulfillment, and so on and so forth. |
Robyn Fadden – host: | Again, every aspect of the organization. In writing your guide to demand prediction in retail, what was your approach to demystifying data science and algorithms for a wider audience that includes retail decision makers, so they could better understand their organization’s needs? |
Maxime Cohen: | Yeah, that’s a great question. I’m teaching several courses at McGill University on AI and analytics. And I’m always looking to find some case studies where we can use real world interesting problems that are data driven, and try to help use data to improve professional decisions. And then when I was talking to many retailers, I kind of got to the conclusion that predicting demand was one of the most important use case. So we tried to write a case study based on demand prediction. And we saw actually, that the case study, to make it end to end that was kind of the goal of this book is to make the entire process from beginning to end from collecting the data, which data to collect, how to aggregate it, how to clean it, how to make sure to visualize the data to understand the data, or the patterns in the data, all the way to evaluate and test the demand prediction accuracy and try to understand what is good material. We really cover everything in the book. It’s more of a technical book, targeted to data scientists and retailers who wants to become more embracing data driven practices. And their will to try and as you say is to demystify how to leverage predictive analytics for demand prediction in retail. And we cover the entire process from beginning to end by going step by step each chapter is covering one other step. And for each step, we illustrate the step by providing the snippets of the programming code that one can do in order to replicate the code. |
Robyn Fadden – host: | Along with the Demand Prediction book, you and your co-authors created a website where there are freely available and downloadable resources, such as a retail dataset, that illustrate the concepts and methods covered in the book. What other resources did you provide and why? |
Maxime Cohen: | The book is coming with a website called demand predictionbook.com. And on the website, we are giving all those sources, we’re giving a data set like data to be used to test all the methods. And we’re also giving all the notebooks which are the codes that need to be used not to replicate. So readers can potentially use their own data from their own store to predict the demand for their own setting of interest. So the website is there with the book. And there was a lot of resources on the website that can be accessible, like free of charge. We wanted to give that and try to democratize access to this type of knowledge, because it can really help a lot of retailers. |
Robyn Fadden – host: | So if someone wasn’t a data scientist, if they were looking through the book, they would see the steps that are involved and the overview of the process. |
Maxime Cohen: | Exactly. There is a small caveat, because if someone is not really like trained in data and statistics, it will be a little bit hard to grasp and apply that book. But at least as you say, they can understand the high level process, what are the steps involved, and they can potentially hire an intern for summer, and give them this material with a data and they can go and explain the processes to make sure to be able to predict the demands. |
Robyn Fadden – host: | That would help them assess goals for this data, for instance. I’m wondering how predictive analytics can be used to optimize retail operations specifically, such as inventory and supply chain management? |
Maxime Cohen: | Going back to the three levels of maturity. So once we’re done with the predictive level, we want to move to prescriptive. So now that we can predict the demand, ideally, accurately, or even very accurately, depending on the data depending on the setting. Now we can say, okay, let’s use this demand prediction, to make the best decision. So now we’re going to prescriptions. So for example, you can decide pricing and promotions, which products to promote at which price discounts and when to schedule the promotions. Now, we can use a demand prediction model with different price points, and try to predict what will be the demand at different price points. And now if we have some specific target in mind, or have a sales target to profit target, we can use a simulation where we use a predictive analytics that we build with a demand prediction. And we can find the optimal price that will satisfy for example, maximizing total profits for the next few weeks. Another use case, as you mentioned, is one of inventory management. Now that we can predict the demand, it can inform the retailers about what is the optimal level of inventory for each product in each week. In each store, they have multiple store and that we are moving again, from predictive to prescriptive once we can predict the demand, we can use these demand prediction values in order to make decisions. |
Robyn Fadden: | So even as you lay out the demand prediction side of the equation and how to implement it, you’re also setting the stage for the prescriptive use of data analytics. |
Maxime Cohen: | Exactly, it’s the way I always like to say it is it is like the foundations. If you have solid foundations, then now you can predict the demand very accurately. Now, we have a lot of opportunities that are offered in order to make sure that you can now do the prescriptions. And at the end of the book, in the conclusion section, we are mentioning all the tools that can be used once you’re done with the predictive in order to go to prescriptive, and again, inventory management, supply chain management, procurement, fulfillments, pricing and promotion, and so on and so forth. |
Robyn Fadden – host: | Now that we’re wrapping our heads around predictive data analytics as a foundation, I’d like to know, what are the main business benefits of using data for prediction purposes? |
Maxime Cohen: | If you leverage data in a smart way you can predict demand accurately. It means that you can do a better job at inventory management because if you know your demand, think about the extreme example when you know exactly perfectly as you might you have a crystal ball. In that case, you’re going to have the exact level of inventory that you need, not one extra and not one less than you need. And then you’re going to have no waste, no stockouts. And that’s perfect. Now predicting the demand. Well, of course, it’s not going to be like a crystal ball perfectly, but it will allow you to reduce waste, which has a huge impact on environments and also to avoid the situation of having stockouts, not having enough inventory to satisfy the demands. We all know that stock outs are very bad for customers reputation for for, for brand reputation, we don’t want customers to come to the store not finding the item of interests. But we also don’t want to have excess inventory. So to mitigate this issue of not having too little inventory, too much inventory, predicting better demand can allow you to do a better job at inventory. And therefore you can reduce waste and increase customers loyalty and satisfaction. Okay. |
Robyn Fadden – host: | It seems like it’s kind of like a no-brainer to decide to use AI in your inventory management, but some organizations aren’t doing that. When do you think it will become the norm? |
Maxime Cohen: | A lot of retailers are using just, you know, like intuition and past experience. And they can do actually a pretty good job sometimes. Now, when you start having a lot of stores and a lot of products, even the most trained managers with decades of experience, will not be able to do as well as very good algorithm to predict the demands. Now, of course, like, there is also a cost: you need to develop those algorithms, to store the data, there is a cost to ingest the data, to clean the data. So there is still a challenge to make sure that even small and medium enterprises that are more local, have modest budgets for infrastructure, and IT, that still cannot afford that’s why we’re trying with this book to democratize a little bit of access by giving away all the codes or the snippets of the notebooks in order to be able to reproduce it at a minimal cost. |
Maxime Cohen: | But I totally understand that there are still challenges. First of all, there was a cultural change. So typically, like change management, and convincing managers who have done their job for many years, to trust a machine or an algorithm to do those type of jobs that were done by humans, there is a big challenge in terms of change management, to convince the managers to be able to adopt those analytics techniques. And the second challenge, as we mentioned, is in terms, of course, like it’s doesn’t come for free. So we need to like you know, recruit data scientists, those are hard to find high salaries, you need also some costs for the data infrastructure, you need some costs for, for making sure that the data is protected in a good environment, there is all privacy issues that you need to make sure that you satisfy in your data. So there is a lot of hurdles and challenges. And that’s why maybe it’s not yet democratize the way I wish it wasn’t. I hope in the future, it’s going to become more and more prevalent. |
Robyn Fadden – host: | And there are some AI analytics tools that are being sold, certainly, but they’re also putting data analytics in the hands of the organization, which could be argued to be part of a wider democratization of access to data and analytics. That brings me to a question of data gathering: we know that most retailers are already collecting data in one form or another from several channels, online and offline. What kind of data is the right kind of data for retailers to predict demand? |
Maxime Cohen: | The first type of data, which is the most common and useful, is point of sales data. So point of sales data are just the transactions. So customer go to the store, and they purchase several products, you need to know the transaction details. So you need to know which products were purchased. What are the characteristics or the attributes of that product, maybe it was a t shirt, what size what color, watch design. Then you have the price, and then you have the timestamp at what time it was purchased. So let’s call just point of sales transaction data that’s the most basic and by far the most important predicaments. The second one can potentially be customer attributes. If you go to the online channel in E commerce, you can know some information about the customer for example, past purchases, for example, if there was some type of login in the system, the customer will potentially have some private information that will be useful for predicting the demands. |
Maxime Cohen: | Another useful data sources is inventory data. If you have access to inventory data that can definitely be also useful to predict the prices, the promotion if there is marketing campaigns, email campaigns or social media. At the end of the day, you need to ask yourself, what are the data that are somewhat correlated or that are helping me to inform demand information, and you can list all the data sources and then you can try to collect as many as possible, of course, feasible and available, and then trying to join all those data sources in order to predict the demand. |
Maxime Cohen: | And last thing I want to say on the topic is there is obviously a lot of internal data from the company. But sometimes now it’s very 20 to also leverage external data. So for example, you can have data from competitors, or some type of aggregation of data per sector, or per quarter for a geographical location. Another example of this data is Google Trends data on Google Trends, you can see the key words, and that can be very useful. A couple of more example, are news articles data, you can look at the words that are mentioned the most commonly in the news articles, recently, that can also help. And the last example will be social media type of data. So good summary is we start with internal data. And the biggest source is point of sales data transactions. Of course, if you can add external, additional internal sources, like email campaigns, promotions, inventory, cost data, and so on, that’s fantastic. And the second type of data is more external. So we gave a couple of examples. For example, competitors, or aggregators, or social media or news articles. Another good, external data can be macroeconomics factor, today, there was a lot of inflation, the inflation can potentially be also accounted as a data to help us predict the demand for specific products. |
Robyn Fadden – host: | Then we bring in machine learning, and it takes that data and analyzes, it connects those marketing campaigns to who’s buying what, on what day, at what time, and what store, for instance? |
Maxime Cohen: | Exactly. You want to build some machine learning algorithms that will be able to get all those data sources as inputs, and then try to understand some hidden patterns of the demand, and try to be able to predict the future demand by leveraging all those data sources from the past. |
Robyn Fadden – host: | And of course, that’s part of the purpose of AI. A human could not do that, or it would take many, many years. |
Maxime Cohen: | It’s even I would say impossible for the human brain to process so much information and to find all the hidden patterns and correlations between different types of features in order to be able to make accurate predictions. And that’s why in that specific case of demand prediction in retail, machine learning algorithms are very useful and have been very successfully applied to get very high prediction demands. |
Robyn Fadden – host: | For data scientists, this book is a goldmine for a specific application of data analysis. How can it help retailers know when they should enter this level of data analysis, when they should hire a data science, or as you said, a summer intern or consultant who can take on this kind of project? How would a company know when the time is right, and is there a way to prepare and set goals for this level of analysis? |
Maxime Cohen: | Unfortunately there is no one size fits all type of answer. My advice is always to start with the pain points. If you’re a retailer, and you feel like your inventory management, or your demand prediction is not good enough, and something you feel like is really affecting your bottom line of business, then you should go and try to solve the problem. I always like to start with a problem and then go to the solution. So if you feel that improving the forecast of your demand will ultimately improve your inventory systems. And that’s something that you are really concerned about, because you believe that it’s not good enough the way it is currently, then that’s probably like a good idea to start at least collecting the data. Now, if you’re already collecting and storing the data, and you’re not using it to predict demand, that’s the right timing. To hire a data scientist, again, it can be some students that come for an internship in the summer with some data analytic expertise. And if the data is available, and they want to potentially use this data to predict demand, that’s something they can do already from day one by using all the material that we share in the book. |
Robyn Fadden – host: | As Maxime Cohen established earlier, a lot if not most of larger retail companies are gathering this data already. The question is how to correlate and leverage data at the lucrative predictive level, what resources to put into those efforts and to what goals. And then, organizations can ask whether they’ll take their data into the prescriptive sphere. But we’ll talk about that on another podcast episode. |
Robyn Fadden – host: | Our guest today on the Delve podcast was Desautels Faculty of Management Professor Maxime Cohen, whose recent book, Demand Prediction in Retail – A Practical Guide to Leverage Data and Predictive Analytics, was published by the Springer Series in Supply Chain Management. You can find out more about the book and its online resources at demandpredictionbook.com. Thank you for listening to the Delve podcast. You can follow DelveMcGill on Facebook, LinkedIn, Twitter and Instagram. And subscribe to the DelveMcGill podcast on your favourite podcasting app. |