00:00:00 Introduction and Sim’s background in data science and supply chain analytics.
00:02:00 Overview of forecasting and time series as a common approach in supply chain management.
00:06:26 Introduction to probabilistic forecasting and its difference from point forecasting.
00:08:10 Sim’s experience with probabilistic forecasting and addressing real-world business challenges.
00:09:15 Discussing the limitations of point forecasting and the benefits of probabilistic forecasting in managing risks.
00:11:39 Discussing quantile regression and its relevance to inventory management.
00:14:20 Transitioning to probabilistic forecasting and its advantages.
00:16:37 Comparing probabilistic forecasting to traditional methods and their limitations.
00:17:37 The development of probabilistic forecasting in different industries.
00:19:00 How probabilistic forecasting simplifies inventory management compared to classic approach.
00:23:07 Discussing the quantile forecasting approach and its challenges.
00:24:26 Constrained environment in businesses and supply chains.
00:25:58 Introduction to probabilistic forecasting for better decision-making.
00:27:56 Addressing supply chain challenges with probabilistic forecasting.
00:30:41 Implementing probabilistic forecasting in large, established companies.
00:34:10 Importance of organizational redesign in adopting a new system.
00:36:02 Traditional approach to conflict resolution between multiple constraints.
00:38:31 Challenges of letting go of traditional forecasting methods.
00:41:00 How to re-prioritize and restructure the resourcing.
00:42:33 The potential value of embracing probabilistic forecasting.
00:45:30 Encouraging businesses to take a leap with new technology.
00:46:55 Transitioning from Excel to more advanced tools in supply chain management.
00:48:30 Dependency on skilled engineers for advanced supply chain tools.
00:50:06 Trusting the recommendations of new systems.
00:53:00 Importance of combined supply chain and data science knowledge.
00:56:49 Creating a cross-departmental data science entity.
00:59:42 Discussing the future structure of data analytics and its incorporation into divisions.
01:02:16 The benefits of having a single data science organization.
01:04:01 The importance of understanding supply chain when implementing new technologies.
01:05:38 Comparing probabilistic forecasting with more complex technologies like deep learning.
01:08:16 Importance of replenishment logic and its relation to forecasting.
01:10:34 The challenges of convincing supply chain leaders to adopt a new approach.
01:12:05 Sim’s experience with probabilistic forecasting and its benefits.
01:15:01 Anecdote of applying quantile forecasting in a large Canadian retailer.
01:19:35 Discussing sub-optimal decisions in business and better decision-making.
01:21:14 The importance of ROI-based decisions for edge cases.
01:22:18 Call to action: embracing new concepts and approaches in supply chain management.
01:23:25 Hopes for a more optimal ordering and replenishment scenario in the future.
Summary
In this interview, Joannes Vermorel, founder of Lokad, and Sim Taylor, Director of Analytics and Data Science at Petco, discuss the importance of probabilistic forecasting in supply chain optimization. They explain that it involves assigning probabilities to multiple scenarios, enabling better inventory management and decision-making. Despite the challenges of implementing probabilistic forecasting in large organizations, they encourage supply chain directors to embrace new concepts and move beyond traditional approaches. They also emphasize the importance of trust in data scientists and the need for professionals who understand both supply chain challenges and technical aspects.
Extended Summary
In this interview, host Conor Doherty speaks with Joannes Vermorel, founder of Lokad, and Sim Taylor, Director of Analytics and Data Science at Petco. They discuss the importance of probabilistic forecasting in supply chain optimization.
Sim Taylor shares his background in data science and analytics, working primarily in merchandise and supply chain analytics. He explains that his work aims to use data and statistics to better guide products to their intended locations in the supply chain to maximize financial results.
When asked about forecasting, Taylor describes it as making an inference or prediction about a future outcome using historical information and other data points. Vermorel agrees, adding that time series forecasting is a cornerstone of classic supply chain theory but emphasizes that it is only one of many approaches to think about the future.
The discussion moves to probabilistic forecasting, which is about assigning probabilities or likelihoods to multiple possible scenarios. Taylor shares his experience using quantile regression in forecasting, which aims to predict extreme events and help prevent stockouts. This method involves solving directly for a specific service level, allowing businesses to set inventory levels without relying on potentially incorrect safety stock calculations.
Taylor explains that the classic approach to inventory management involves predicting the most likely outcome (forecast) and accounting for variability by adding safety stock. Probabilistic forecasting, on the other hand, directly estimates the total inventory needed to cover the risk of stockouts. Taylor also talks about the challenges of determining the right quantile extreme of demand and balancing customer needs with financials and inventory costs.
The interviewees discuss the advantages of probabilistic forecasting in constrained environments, such as limited budgets or minimum order quantities. Taylor notes that probabilistic forecasting can help businesses make clear decisions between competing needs by understanding the likelihood of selling each additional unit and the financial value it would provide.
The discussion highlights the challenges organizations face when trying to implement probabilistic forecasting, especially when it comes to altering established processes and divisions of labor. Sim Taylor shares his experience in implementing the quantum forecasting approach, which requires a comprehensive understanding and ownership of supply chain execution. The forecast serves as a means to an end, focusing on improving customer satisfaction, sales, and minimizing costs.
Sim Taylor stresses the importance of automating the mathematical heavy lifting in forecasting and order quantities, while still acknowledging the value of human intuition and expertise from experienced supply chain managers. He suggests that businesses can redesign their organizations to optimize the use of their experienced staff, which may involve changing their roles and relying on the system’s recommendations.
Joannes Vermorel emphasizes the challenge of letting go of traditional forecasts and their associated KPIs, as probabilistic forecasting shifts the focus away from the accuracy of the forecasts and towards the operational decisions enabled by them. He asks Sim Taylor for his perspective on how businesses can re-prioritize and restructure their resources to accommodate this change.
Sim Taylor acknowledges the difficulty in implementing probabilistic forecasting in large, risk-averse organizations. He suggests that success can be achieved step by step, by first delivering simpler and faster tools, then demonstrating the value of the new approach. This may involve finding a use case or company willing to take the leap of faith and measure the business results. Overall, the conversation emphasizes the potential benefits of adopting probabilistic forecasting, while recognizing the challenges it presents for established organizations.
The conversation touches on the need to demonstrate success to encourage adoption of new technologies and approaches. Vermorel raises concerns about the limitations of Excel and the need for more advanced tools to handle uncertainty. Taylor acknowledges the use of Excel by supply chain practitioners but highlights that many decisions are driven by a system that is often a “black box” to those using it.
Both Taylor and Vermorel emphasize the importance of trust in data scientists and the need for professionals who understand supply chain challenges and can work with code to automate decisions intelligently. Taylor suggests finding individuals with a combination of supply chain understanding and technical inclination. Vermorel adds that in the future, he envisions data science being integrated into each division of a company, with analysts specializing in their respective areas.
Sim Taylor emphasizes the importance of delivering tangible value in supply chain management, regardless of one’s title or role. He believes that combining technical expertise with business knowledge is crucial for achieving positive financial outcomes. The conversation also touches on the value of collaboration between analytics experts from different areas, and the importance of being tied to the business for meaningful results.
Joannes Vermorel highlights the practical benefits of probabilistic forecasting in supply chain management, as opposed to more technical innovations like deep learning. Probabilistic forecasting focuses on creating value by approaching the problem of uncertainty in a radically different way.
Sim Taylor also discusses the importance of replenishment logic and how it is crucial for making decisions in supply chain management. He believes that focusing on supply chain value is essential, and the conversation should center around how decisions drive the best output.
Both Taylor and Vermorel agree that sharing use cases and success stories is important for building trust with businesses and demonstrating the value of probabilistic forecasting. Taylor shares an example of working with a large Canadian retailer, where the application of a quantile forecasting approach led to improvements in stock availability and inventory management.
Taylor emphasizes that supply chain is a world full of constraints, and that edge cases can lead to sub-optimal results. The goal is to use additional information to inform more optimal decisions that require less effort. An example he provides is deciding whether to order a full truckload from a vendor when only a portion is needed. The ability to assess the value of each unit, the likelihood of selling it, and the potential margin and holding costs enables better decision-making in such cases.
Both Vermorel and Taylor encourage supply chain directors to evaluate and embrace new concepts, such as probabilistic forecasting, and to move beyond traditional approaches that have been used for decades. They hope that by sharing successful examples, they can accelerate the conversation and help businesses make more optimal ordering and replenishment decisions.
Full Transcript
Conor Doherty: Welcome back to Lokad TV. I’m your host, Conor, and as always, I’m joined by Lokad founder, Joannes Vermorel. Today, we’re talking to Sim Taylor. Now, he’s the Director of Analytics and Data Science at Petco, a pretty big company, and he’s going to be talking to us about the merits of probabilistic forecasting. Tim, welcome to Lokad.
Sim Taylor: Hi, thanks very much for having me, Conor and Joannes. So, a little bit of background about myself: I manage data science and analytics teams, specializing in merchandise and supply chain analytics. A majority of my career has been from a consulting perspective. I worked with Deloitte for many years in their supply chain teams in the UK and Canada, and really in the latter half of that, I focused on building models to help businesses optimize their inventory. Essentially, the work that I do boils down to how we can use data and statistics to better guide products to various locations in the supply chain, fulfillment centers, stores, customers in the right quantities at the right time to satisfy customers and maximize financial results. I’ve worked with a range of companies, predominantly department stores and general merchandise retailers, typically with a lot of products, a big range, and a large stock portfolio with some fashion and specialty retail thrown into the mix. A couple of years back, I transitioned from the consulting world into industry, and I’m currently working at Petco, a large American pet retailer based in California with a presence across the US and Mexico.
Conor Doherty: Thank you. When you describe to people just generally, or in general terms, when you describe to people what forecasting is, how do you explain that to non-specialist audiences?
Sim Taylor: Sure, so forecasting, as a baseline, is a relatively understandable process. It’s essentially making an inference or prediction about a future outcome. In supply chain and retail, we talk a lot about forecasting demand, customer demand. So, for this product at this store, what do my customers want one week from now, or two weeks from now, or 52 weeks from now? It’s really trying to use historic information and things we know about the future, like the prices of products and other key data points, to understand what is a likely outcome that we could expect to see and how can we plan our ordering and getting the product in the right place around that. It’s really the cornerstone of classic supply chain theory, that forecasting. Joannes, would you agree with that definition?
Joannes Vermorel: Yes, I very much agree with the idea of time series forecasting as the cornerstone of classic supply chain theory. But also, I would like to point out that it is actually a very specific flavor of the way you can approach the future. This approach has been so established and so prevalent for so long that the industry as a whole has, to a large extent, forgotten that it is only one of the many approaches that exist to think about the future. Because it’s an elusive thing, thinking about the future and anticipating it. To think about it through the lenses of time series forecast gives you some capabilities, like plenty of models that would fit into this paradigm, but also it restricts what you can do. And I think one of the interesting things in this industry is that due to the fact that time series forecasts have been around
Conor Doherty: I think it might be helpful to talk about how I came across probabilistic forecasting. I’m from a business background rather than academic, so I actually came across probabilistic forecasting when working with a retailer trying to solve some of the real-world business challenges of ordering and replenishment.
Sim Taylor: For that customer, we were building replenishment tools based on the standard point forecast, which is still the baseline across most businesses. Essentially, we think that this product is going to sell five units tomorrow, six units the day after, etc. Because there’s so much possibility for that one number to not be right, we were then applying the classic safety stock approach or equation, basically saying that our single prediction of the most likely demand outcome is going to be wrong a lot of the time.
The standard safety stock approach essentially makes a blunt assumption that the variability of demand and supplier lead times are symmetrically distributed around our most likely forecast prediction. This assumption of normality from a stats perspective really just gets accepted as the baseline without question at a lot of businesses.
As we were doing this work, my team and I were looking at the actual reality of sales across the relevant lead time window. We came across the insight that, in the majority of cases, the assumption of normality does not hold at all for most retailers. Demand and lead times are not typically normally distributed. Your demand tends to be highly consolidated around a low value for most products, and then there’s this long tail of potential demand values out to the right.
It became clear to us that we needed to use more representative statistical distributions to model this variability of customer demand and lead time. Otherwise, we’d be at risk of stocking out in some situations because we haven’t modeled the potential outcomes correctly.
Conor Doherty: So how did you and your team start considering alternatives to the point forecast?
Sim Taylor: We first ended up leaning on quantile regression in our forecasts, which centers on the insight that, in inventory, we don’t really care about the most likely outcome, which is what we get from a standard forecast. What we care about is those extreme events and making sure that we have enough inventory in place to prevent stockouts when these happen. That’s what service levels do in the standard safety stock approach.
We built forecasts using quantile regression that tried to solve directly for a specific service level. For example, if we list out all the possible demand occurrences across the lead time, what’s the 95th quantile, the very unlikely situation of that demand in that lead time? That’s what you set your inventory level at. You don’t have to worry about calling it a safety stock or if your safety stock calculations are inherently wrong based on incorrect assumptions.
That’s how we started exploring different demand forecasts other than just the mean or the average outcome. Then, the question evolved to what should we set our service levels at to determine what quantile of demand or lead time to predict, and how do we combine those quantile forecasts for demand and lead times? That’s when I came across Lokad and their unique, differentiated approach where they forecast all possible outcomes for demand across the lead time and assign a probability to each, which captures far more information than even understanding the quantiles.
Joannes Vermorel: Your journey is incredibly reminiscent of mine, because it’s at Lokad that I had my personal problems with these assumptions. The term “safety stock” sounds safe, but the reality is that the mathematical assumptions that go into it are pretty insane and certainly very much unsafe. You end up with things like negative lead times, which is incredibly bizarre. That’s what you get with a normal distribution for lead times. And the idea
Conor Doherty: I realized that there was a mismatch between what people were truly interested in and what they were saying. Supply chain directors wanted a better forecast, but what they ultimately wanted was a better supply chain decision. When you start thinking in terms of statistics, it’s not the average that matters, but the extremes. These extreme situations are what trigger out-of-stock or excess stock scenarios, and those are the situations that you truly want to analyze from a statistical perspective. I found that people in finance and weather forecasting have been doing probabilistic forecasting since the early 90s. In 2011-2012, we started doing that for supply chain with quite a lag compared to other pioneers in other verticals.
Sim Taylor: The classic approach still tries to account for risk, but we’re just calling our inventory different things. In reality, what matters is the outcome to the business. In our standard approach, we try to get really good at predicting the most likely outcome, but we acknowledge that it’s a strange way of looking at things because we really care about covering for the risk of more extreme eventualities, and we have that safety stock. Quantile forecasting is directly setting both of those, looking at the extreme example of demand and calculating the inventory we should have in our location.
Joannes Vermorel: When you have a full range of potential likelihoods, how does that actually get transformed and used in business reality? Those ordering decisions are what we care about, making sure that they are as optimal as possible. We might not need the term “safety stock” anymore, but I’m curious about how you translate that into real-world execution.
Sim Taylor: The first motivation to consider alternative forecasts was seeing the reality that demand occurrences don’t follow the type of distribution that is modeled in the classic safety stock approach. That was the main impetus for me to consider other approaches and evaluate our ability to have an alternative model that better fits the data.
Conor Doherty: So, to better replicate or account for variability, as it typically looks for demand and lead times, you started with a quantile forecasting approach. But there were challenges, like determining the right quantile extreme of demand to balance customer needs and financials versus the cost of holding inventory. And in constrained environments, which are common in businesses and supply chains, you needed a way to decide between different products and units to purchase. Can you talk more about how you came across probabilistic forecasting as a solution to these challenges?
Sim Taylor: Yes, what interested me in probabilistic forecasting was the concept of understanding all the potential scenarios and their likelihoods. If I know the likelihood of selling four units, five units, six units, I can work out how likely I am to sell the next unit I’m going to purchase. If I know that and the value I get from selling that product, the gross margin, as well as the costs of holding that product and stocking out, I can make a clear decision about which unit provides more value to my business. In doing that, it solves many challenges that my team and others have been grappling with for a long time. We’re always in a constrained environment, and understanding the likelihood of each next unit selling and its financial value is an elegant way to address those challenges.
Joannes Vermorel: That’s very interesting. My biggest struggle when I try to push probabilistic forecasting to large established companies like Petco is the historical division of labor between forecasting and decision-making. Large organizations have typically divided the job between people in charge of forecasting and those making decisions like replenishment, production orders, and stock movement orders. But probabilistic forecasting is a tool that enables a more efficient decision-making process. The decision-making itself is not exactly the forecast, but fundamentally, taking the replenishment decision and making the forecast are much more entangled than in the old world where there was a separate team doing safety stock analysis and making the decisions.
Conor Doherty: Replenishment is, I would say, a classic supply chain theory that provides a very clean division of labor between the people who are in charge of doing the forecasting/planning and then the other people who are supposed to take the operational decision, like production orders, replenishment orders, stock movement orders, etc. How do you approach that? You’ve operated in very large, very established companies that obviously had a supply chain practice that largely predates those probabilistic insights. How did you actually approach the organization with this sort of method that, I believe, does not respect the historical boundaries of the division of labor that existed to support the supply chain practice?
Sim Taylor: I think in situations where we’ve returned value from, for example, the quantum forecasting approach, it really helps to have ownership across the entire supply chain execution. If our ultimate goal, which it always is, is to improve customer satisfaction, sales, and minimize our cost while doing so, the forecast itself is just a means to an end. We care about the outcomes: how much should I order, to what location, and at what time, both in distribution centers and in stores. Where I’ve had success before is when we were given the remit of constructing that forecast and directly translating it to the ordering quantities.
We forecast, generate the order quantities, either from vendors into distribution centers or replenishment from distribution centers into stores. The role is very much an approval and evaluation process. In an ideal world, you want to automate as much of the mathematical heavy lifting and get things right across the majority of scenarios. Then, combining that data-driven insight with the expertise of business teams is a way of understanding and evaluating the extremes or various examples and making adjustments there.
When a business is looking to replace or transform the way they do their ordering replenishment, it’s a sizable project that requires a lot of work, many months to implement, and a lot of people’s involvement. It’s a chance to reset and refresh how your business runs. Many businesses use that as an opportunity to redesign their organization, which typically involves change management and organizational redesign. If you can redesign your organization and reframe it around a much better system, that’s where the real success comes in reducing ordering costs and improving stock levels for customers.
Conor Doherty: I just have a very specific question that I want to throw to you, Joannes. We talk about the traditional approach to resolving the conflicts between multiple constraints when we talk about business teams. What’s the normal way that a non-probabilistic business team would resolve that, or a non-probabilistic optimization process? How is that done on a day-to-day basis?
Joannes Vermorel: If we take a retail example, you would have a team that does forecasting. They would establish the baseline, manage seasonality profiles, and maybe establish things like ABC classes with the level of focus and whatnot. So there would be a team very much related to establishing projections in the future that are time series. Then you end up with another team or a series of other teams that are responsible for the operational decisions, like replenishment or deciding whether a product is even eligible or not in a store. The mainstream organization is fully justified, and it’s very established and rarified in the sense that most of the software tooling that exists in the market gives you user interfaces and processes.
Joannes Vermorel: When you have tailored workflows for this mindset, even if it’s partly arbitrary, once you have tons of tools that verify the positions, you have a screen dedicated to the review of time series forecasts, another screen for the adjustment of safety stocks, and another screen for alert management. These abstract concepts become tangible things in the organization because there are people with roles and workflows built on top of that.
As our guest mentioned, an interesting thing is that probabilistic forecasting challenges the organization at a fairly deep level. If we have a sizable project that revisits one of the core functions of the company, such as replenishment in a retail business, there is an opportunity to revisit many assumptions. It takes a bit of a leap of faith, as it is a leap from the traditional way to approach supply chain management to an alternative organization that is a consequence of what probabilistic forecasts let you do in your company.
Forecasts are just artifacts; they are an instrument to get something else. And yet, in many companies, especially those with a large supply chain, forecasts are typically treated as a goal in itself. There is a team with KPIs in accuracy, and this is part of the S&OP process. They want a more accurate forecast as a goal for the next year.
Sim Taylor: The challenge is to let go of the forecast. With probabilistic forecasting, you’re proposing to shift the focus of management. Instead of following forecasting accuracy KPIs, providing resources, budget, and tools to improve them, we should focus on the decisions we make based on these forecasts. This is a lot to take in, and it requires a leap of faith. How can we give a deeper argumentation rather than just saying “trust me, it’s going to be so much better”?
It’s hard and very challenging. I see the huge potential value here, but I don’t know of any large business that has fully embraced probabilistic forecasting. Implementing it step by step, delivering a simpler, faster tool first, and showing results is the way to engage and infuse supply chain executives. You need that one example, and then measure the business reality: the in-stock availability, inventory turns, inventory investment, and weeks of supply before and after rolling out the new approach.
Show the clear, undeniable financial benefits driven by the new approach. Do we really care about the forecast? We can talk about that, but the outcomes are what matter. Leading with the business outcomes, and starting the conversation with historical results and success will pique the attention more than leading with theory.
Conor Doherty: I mentioned supply chain experts, leaders with lots of experience, they have this great intuition, and you can tap into that. Everyone wants to improve their business, so almost everyone has an impetus to want to take a leap, but they just need some grounding and some faith. They’ve seen a similar company and the financial returns that really demonstrate success. That’s how, you know, once you’ve had that one use case of taking the leap, you then really have a platform and can start to roll that out. I think that’s how new technologies and approaches always start. At the end of the day, we’re there in a business to do the best by the customer and to do the best by shareholders financially. If we can show that there’s a good use case to take that risk, try something different because we’ve seen returns elsewhere, that’s an exciting proposition. You’re much more likely to start having a conversation and get the ball rolling. But I do agree, it’s a big challenge because established processes are well-rooted in the way that we structure our teams across most large businesses. As a result, it takes the right people who are willing to really understand the challenges and can also influence to make that business case, to show there’s actually value in trying something like this.
Joannes Vermorel: One of the objections I frequently get is that when you transition from something like point forecasts, which can be represented easily in Excel, to probabilistic forecasts, which can be represented in Excel to some extent, it becomes a bit of a nightmarish process. So, the bottom line is that you need to upgrade your tooling one way or another. In order to be able to have models that embrace uncertainty, you need a tool that is more capable than Excel. You create a class of risk, which is suddenly your organization depending on people that are capable of wielding those more complex tools. Due to your position as head of data science, I think you’re on the forefront of introducing stuff that is not immediately accessible. As soon as we are talking about having 20 lines of Python, the sort of skill needed to take care of that, as opposed to having just a simple spreadsheet, is quite large. So which means that for the company, you’re dependent on people that have a lot more engineering skill than what they used to have in those supply chain positions. I know that some of the people I’m talking to are worried about this sort of proposition. They wonder how to address that when they want to make a real-world supply chain dependent on something that is not going to fly through Excel spreadsheets or be implemented as a rule-based system in ERP.
Sim Taylor: There are probably two things to unpack. The first is that Excel is a tool that’s very familiar to category managers, buyers, and supply chain practitioners. There’s definitely use there, but in most of the companies, the majority of the decisions are being determined and driven by a system that is already kind of a black box for those people who are actually making the purchasing decisions. So, the point you raise about moving from Excel isn’t a huge leap in some cases. It’s a transition from one system that makes recommendations in one way to another system that makes recommendations in a different way. I think as long as there’s trust in how the new system works, and the right education and discussions to understand why the system is doing something slightly different, it’s not so much of a leap to change the system itself.
The second part is the trust in data scientists. I do really believe that supply chain and merchandise planning is one of those areas you can’t just throw a data scientist at and expect tangible financial benefits. There’s perhaps a
Conor Doherty: Joannes and Sim, I want to know more about the importance of having both technical and mathematical expertise as well as supply chain experience when it comes to optimizing supply chains. What are your thoughts on this?
Sim Taylor: Supply chain optimization requires a combination of expertise, including understanding the challenges and complexities of the supply chain world as well as having mathematical and technical abilities. The ideal candidate would have experience in both supply chain management and data science. They should be able to understand and work with previous systems and have knowledge of common challenges, such as vendor minimum order quantities and capacity constraints at distribution centers. A data scientist straight out of grad school might not have this knowledge, which can lead to distrust when trying to marry technology with supply chain management. What we’re looking for are smart individuals with both business understanding and technical inclination, capable of working with or explaining code to automate decisions in a smart way. That combination is crucial for building successful analytics teams.
Conor Doherty: Joannes, would you say that your approach to selecting supply chain scientists at Lokad aligns with Sim’s views on the importance of having a combination of skills?
Joannes Vermorel: Yes, I think our views align. In fact, I see a possible evolution in the market where, a decade from now, data science teams that currently operate independently might become part of supply chain divisions. I envision a future where each division has a team of engineers specializing in quantitative analysis, providing the backbone of optimization for the company. They would work closely with operations and top management to establish strategies. I believe this natural evolution would lead to the reincorporation of data science skills into each respective department, such as supply chain analysts becoming part of the supply chain division, and marketing analysts joining the marketing division. The current organizational structure, where data science departments are independent from the divisions they serve, might change as data science becomes an intrinsic element of each practice rather than a supporting function.
Sim Taylor: Joannes, I think your perspective on the future of data science and analytics within organizations is interesting. As the head of analytics at Petco, I believe that if we continue to see operational successes, there may indeed be a shift towards incorporating data analytics divisions into their respective departments a decade from now.
Conor Doherty: Do you see that as more of the things being internalized in every division where you do achieve success?
Joannes Vermorel: What an interesting question, really talking about how you organize and structure a business now and what the benefits of those different approaches are. From my personal opinion, I don’t disagree with you. I think the title and the organization you’re in is somewhat irrelevant.
Sim Taylor: What excites me when I go to work each day is trying to find opportunities to deliver value, tangible value that we can measure. I focus very much specifically in the supply chain, as I mentioned, like in merch planning. My boss runs the broader data analytics org across many different areas and kind of draws those together. But as I mentioned, I think specifically for supply chain, I’ve seen this in the past: you have really smart folks who don’t have the supply chain experience but are very technical, and they don’t necessarily equate to delivering that meaningful value in the business. It’s the combination of the business skill set and really understanding that industry with the ability to execute technically that I think drives positive, meaningful financial outcomes.
So whether you call yourself a data scientist or you call yourself a supply chain practitioner, in some ways, I think it is the wrong question or an irrelevant question. For myself, I get excitement from solving supply chain problems, and so whether you call folks in a team supply chain scientists or demand planners, what matters is the work you do.
I definitely agree that you need to be very tied in with the business; otherwise, there’s so much nuance in the world of supply chain and merch planning that if you don’t have a grasp on that, you’re just playing around with numbers, and it’s hard to deliver some meaningful results.
The one flip side is, I think there’s a lot of value that can come from grouping up analytics experts and practitioners from different areas. The benefit of having a single data science org is that you can build those connections and relationships between like-minded, smart folks, and they can learn from each other and really draw on some of your technical expertise from different areas. There’s a lot of benefit in that and how the teams interact together across marketing analytics, supply chain, customer, pricing, for example. There’s a lot of connection that can be valuable and bring that together.
I think it’s really down to the company and what they see as the overall best strategy. But just purely from a supply chain perspective, I think the results are what matters, and I think it’s very hard to get that trust that we talked about, which is so important in trying something new and seemingly complex, if you can’t distill it and explain it in terms of how it is going to drive business value with an understanding of how supply chain works.
Joannes Vermorel: I think you touched on a point that is very interesting, and I’m backtracking a little bit compared to your comments about centralization. To go back to probabilistic forecasting, it is a fairly technical innovation, but where it’s surprising is that it touches so much the business. The reason why it interests you, and you’ve mentioned that quite a few times, is the values that you can create.
The interesting thing from my perspective, and that’s also something that is very difficult to convey, is that it is something in nature that is incredibly different from, let’s say, deep learning or AI. This is fundamentally an innovation that is of interest due to what you can do with it in a very practical supply chain way. The amount of technological ingredients, there are some technological ingredients into doing probabilistic forecasting, but it plays a
Conor Doherty: Better forecasts can lead to an improved supply chain. Companies that excel in the classical approach can have better ordering and replenishment optimization. However, it’s important to note that the replenishment logic and how we extrapolate forecasts into supply chain decisions is equally important.
Sim Taylor: People are often surprised to know that in some cases, a naive rolling average could provide similar results to a poor point forecast. I try to focus the conversation more on making decisions based on what will drive the best output, and how we evaluate demand variability and lead time variability. We need to demonstrate the financial success of this approach in a simple and tangible way to businesses.
Conor Doherty: Sim and Joannes, do you have examples from companies or clients where you can truly show the financial improvement after implementing this approach?
Sim Taylor: I don’t have a specific example from a probabilistic forecast perspective, but from a quantile perspective, I have seen success. We rolled this approach out with businesses, focusing on key business metrics like in-stock availability and weeks of supply. We were able to achieve higher in-stock availability while keeping inventory levels the same or even reducing them. This approach helped build trust and showed that intelligently applying these concepts can drive business value.
Conor Doherty: Do you have any anecdotes of success stories where you have applied a probabilistic forecasting approach and demonstrated the value that you talk about?
Sim Taylor: Not specifically from a probabilistic forecasting perspective, but from a quantile perspective, I do. Working with a large Canadian retailer, we rolled out a point forecast safety stock approach, which led to some improvement in in-stock availability without increasing inventory. When we moved towards quantile forecasting, we saw a meaningful growth in achieving higher in-stock availability while keeping inventory levels the same or even reducing them. This showed the clear trends and differences in business performance factors, building trust in this approach. I am an enthusiast who wants to take the next leap and use the probabilistic forecasting approach to further improve these results.
Conor Doherty: Quantile forecasting doesn’t quite solve, well, thank you, Sim, on that point. To throw it over to you, Joannes, any anecdotes like Sim’s?
Joannes Vermorel: Yeah, I mean, it’s a detail, but I think if I compare what Lokad was doing a decade and a half ago when it started, and now what it’s doing with probabilistic forecasting-enabled approaches for supply chain, the amount of edge cases is very interesting. When we started, there was like every single time we were trying to address a supply chain problem, a forest of edge cases. By edge cases, I mean something where the usual logic just failed so badly that you need to step in and manually correct the thing because the result is nonsensical. That’s the reason why vendors like Lokad typically have alerts and exceptions to manage all those situations that are so blatantly nonsensical. You can even have a rule to detect it and say, “Okay, somebody needs to step in because the system is producing an output that makes so little sense.”
And we went from having tons of it to very little of it. So it’s interesting because, for me, that has been the elimination of the edge cases, the sort of things that need specific attention, specific rules. In the end, I believe that this elimination proves that all of that was just the consequence of having a method that, at its core, ignored uncertainty. Thus, whenever there was uncertainty or risk kicking in, and by the way, when we have constraints, for example, with an emoji, and whatever it means to make a risk-aware decision, predominantly, we managed to simplify them.
Although it’s a very anecdotal element, eliminating edge cases, I have found that it was the most reliable indicator of the maturity and quality of your technology. It is whether you can operate with a very limited amount of edge cases and fringe situations that require human intervention and micromanagement of the software solution. Conversely, if you have to micromanage your software solution that operates your supply chain, most likely, it means that at the fundamental level, there is something you’re not getting right. You have an impedance mismatch in the way you approach the problem and the way your software operates, and thus you end up having this micromanagement as a way to duct tape your supply chain.
For me, the question will be, can we move from point forecasts to quantile to probabilistic forecasts? So the open question will be, what’s next? I’m pretty sure that there will be a next stage of technology in this area, but for now, the battle still remains to even get most companies to come to terms with the idea of just accepting uncertainty.
Sim Taylor: And Joannes, just to add, you know, you talk about the work reduction as a major benefit that you’ve seen. I think the flip side is, if we talk about some of those niche edge cases, which really aren’t so edge, they just crop up all the time in supply chain, as we said, it’s a world full of constraints. Either it takes more work, or what I see in previous businesses, much more often, the wrong decision gets made. Maybe it’s not extra work, but we just take a
Conor Doherty: Simplified decision-making often leads to sub-optimal results. The goal is to use additional information to inform a more optimal decision that takes less effort.
Sim Taylor: Right. For example, we only have enough purchasing needs to order 50 truckloads from a vendor, but our contract stipulates we must order full truckloads or meet a minimum order quantity. Is it financially better to order now and buy additional product we don’t need, leading to excess inventory taking up capacity, or do we not order at all and risk stockouts on certain products? Ideally, we’d have an ROI-based decision on whether we should order now or later, which would make edge case decisions more effective.
Conor Doherty: Gentlemen, I don’t have any other questions. Sim, as a customer, I’ll give you the last word. Is there anything you’d like to add or any follow-up questions you’d like to pose to supply chain directors who haven’t yet embraced probabilistic forecasting?
Sim Taylor: My call to action for the supply chain community is to evaluate these concepts, embrace them, and take the leap to test and iterate to understand the value we can gain from moving beyond the classic approach. Are we just replacing one system with another that does a similar thing? I’d love to see more conversations and real use cases where we start to think about this approach and embed it in businesses to get success from it. As we share examples of the value, it’ll help to accelerate the conversation and move us collectively towards a more optimal ordering and replenishment scenario.
Conor Doherty: Thank you for that, Sim. I can’t take all the credit. Joannes helps a lot.
Joannes Vermorel: Thank you, Conor. I appreciate it.
Conor Doherty: On that note, gentlemen, we’ll call things to a close. Joannes, thank you very much for your time. Sim, thank you very much for yours. We’ll see you next time.
