00:00:07 Introduction to Stefan de Kok and embracing uncertainty in supply chains.
00:00:34 Stefan’s background, founding of Wahupa, and initial struggles.
00:03:18 Different classes of uncertainty in supply chains and their impact.
00:04:53 Lokad’s approach to dealing with uncertainty using probabilistic forecasting.
00:06:43 Traditional approaches to handling uncertainty: buffers, response mechanisms, and ignoring.
00:08:01 Consequences of ignoring customer needs and relying on response.
00:09:59 Discussing the journey to probabilistic forecasting.
00:12:35 Stefan’s epiphanies and embracing probabilistic forecasting.
00:14:01 Realizing the importance of the concept rather than the method.
00:15:25 Importance of changing traditional metrics for better forecasting.
00:18:38 Helping customers embrace uncertainty and strategies to employ from a software perspective.
00:20:19 The impact of a company changing to a probabilistic approach.
00:22:38 Observing the market’s readiness to accept uncertainty and what excites the interviewees for the future.
00:25:01 Closing remarks.
In a Lokad TV episode, host Kieran Chandler interviews Joannes Vermorel, founder of Lokad, and Stefan de Kok, co-founder and CEO of Wahupa, discussing supply chain uncertainty. They emphasize embracing uncertainty and considering all potential outcomes for better management. The traditional method of scenario planning is resource-intensive, while probabilistic forecasting offers a concise solution. De Kok identifies three ways companies handle uncertainty: using buffers, responding to it, or ignoring it. Both guests advocate for adopting probabilistic forecasting, being open about methods without revealing the “secret sauce,” and utilizing probabilistic metrics for decision-making. They foresee mainstream adoption of probabilistic approaches in the future.
In this episode of Lokad TV, host Kieran Chandler interviews Joannes Vermorel, founder of Lokad, a software company specializing in supply chain optimization, and Stefan de Kok, co-founder and CEO of Wahupa. The discussion focuses on embracing uncertainty within supply chains, which is traditionally managed using buffer stock. The guests also discuss their backgrounds and the companies they founded.
Stefan de Kok, a co-founder of Wahupa, began his career in applied mathematics at the Technical University of Delft in the Netherlands. After entering the job market, he accidentally encountered a supply chain software company and joined them. Over time, he worked in various roles, such as consulting, product management, and functional consulting. After finding himself out of a job, he decided to act on his ideas and address the issues he had discovered over the years. His initial idea was to create a platform that would be available to smaller companies, as existing products were primarily targeted at large Tier 1 companies. In 2003, he struggled to find people capable of building the platform, but eventually found a team to bring his vision to life.
Uncertainty is a core aspect of supply chain management, as everything in the future is potentially uncertain. Some examples of uncertainty include lead times, durations, quality or grade yields, and rates. Stefan believes that supply chain professionals should consider the impact of all possible combinations of future outcomes, even though it may be a complex task.
Joannes Vermorel, the founder of Lokad, shares his thoughts on approaching the challenges presented by uncertainty. Traditionally, companies used “what-if” scenarios to prepare for uncertainty, but this approach can quickly become tedious. In order to effectively manage uncertainty, supply chain professionals must consider all possible futures and their potential impacts.
Vermorel shares that the traditional method of using scenarios to deal with complex supply chain problems is both time-consuming and resource-intensive. However, probabilistic forecasting offers an elegant, concise solution that can be implemented using raw processing power. This approach has the added benefit of requiring fewer people to manage and operate a supply chain, making it more efficient both from a software and operational standpoint.
De Kok explains that there are three main ways companies deal with uncertainty in supply chains: by using buffers, responding to uncertainty as it happens, or simply ignoring it. Most companies employ a combination of these approaches, but the challenge lies in striking the right balance between them. If buffers are not accurate, companies need to overcompensate by responding, which can be expensive. The aspects that cannot be responded to are often ignored, which can lead to long-term damage, customer dissatisfaction, and potentially bankruptcy.
De Kok also highlights the role of buffers, such as lead times, capacities, and inventories, in managing supply chain uncertainty. Companies often have inflated buffers in order to prevent issues that require responsive action. However, he points out that many companies still struggle to achieve their targeted service levels, as their actual performance is often driven by the responsive actions rather than the buffers.
Vermorel explains that his company, Lokad, started with classic forecasting but eventually transitioned to probabilistic forecasting. They initially used quantile forecasts, which intentionally introduce a bias to account for situations where forecasting the mean would be inaccurate. They then progressed to using quantile grids, which involved incrementally increasing biases, and finally to probabilistic forecasting, which takes into account all biases at once.
De Kok shares that he had an epiphany regarding the value of probabilistic forecasting as early as 2006, when he realized that uncertain values cannot be represented by exact numbers. He began developing a probabilistic arithmetic and found it to be an elegant solution to complex supply chain problems. Initially, de Kok kept his probabilistic forecasting approach a secret, as he considered it a key differentiator for his business. However, he eventually discovered that other companies, including Lokad, were also using similar methods, proving the viability and value of probabilistic forecasting in the industry.
Stefan de Kok emphasizes that there are various ways to achieve the same goal in supply chain optimization. He highlights four key points: 1) embracing uncertainty; 2) adopting probabilistic forecasting and planning; 3) being open about the methods and ideas without revealing the “secret sauce”; and 4) recognizing that traditional metrics are insufficient and need to be replaced by probabilistic metrics. Both Vermorel and de Kok agree that changing the metrics and utilizing probabilistic approaches are crucial for better decision-making in supply chain management.
Vermorel explains that once a probabilistic model is in place, it allows for the simulation of multiple possible futures, which in turn enables the assessment of decisions and their potential outcomes. He also notes that the technology and algorithms used for forecasting are less important than the overall approach, as the evolution of Lokad’s own forecasting engines demonstrates.
De Kok asserts that to help customers embrace uncertainty, it is essential to provide self-explanatory outputs that inspire trust. Visualization of results is critical, as it allows users to understand the range of possible values and their probabilities. He likens the relationship between software providers and customers to mechanics and drivers, with the former creating sophisticated tools for the latter to use easily and effectively.
When discussing the impact of a probabilistic approach on businesses, Vermorel points out that it often provides insights that align with intuition. For example, a classic forecasting method may suggest overstocking perishable goods, while a probabilistic approach would more accurately balance the risks associated with stockouts and spoilage.
In terms of market acceptance, de Kok observes that there has been pushback against embracing uncertainty, but that resistance is gradually decreasing. He identifies two stages of acceptance: first, recognizing that exact numbers are not sufficient for dealing with uncertainty, and second, overcoming misconceptions about the complexity of probabilistic approaches. He expresses optimism about the growing trend towards embracing these methods and anticipates that they will eventually become mainstream in the industry.
The interview highlights the value of embracing uncertainty, adopting probabilistic metrics, and utilizing visualization for better decision-making in supply chain optimization. Both Vermorel and de Kok advocate for the continued evolution of forecasting technology and foresee a future where the mainstream adopts probabilistic approaches.
Kieran Chandler: Today on Lokad TV, we’re delighted to be joined by Stefan de Kok, the founder of Wahupa, who’s going to explain to us why this uncertainty should not actually be seen as a hindrance but something that should be embraced. So, Stefan, thanks very much for joining us today. Perhaps you could start off by telling us a little bit about your background and also Wahupa, the company you founded.
Stefan de Kok: Well, thank you, Kieran, and Joannes for having me. Yes, so I’m one of the co-founders of Wahupa. I started studying applied mathematics at the Technical University of Delft in the Netherlands and never heard of this thing called supply chain. After I hit the job market, by accident, I encountered a supply chain software company and joined them. I never regretted it for a single moment. Then, I did a lot of work for them and a lot of customers in consulting, software consulting, functional consulting, software product management. Then, after another chance encounter, I found myself out of a job, and that’s the moment when I realized that all these ideas I’d been working on, all the issues I discovered over the years, I could actually do something about them now, things that I wasn’t able to do before.
The original idea was to build a platform that was something bigger than S&OP at the time, which included a lot of the best-of-breed solutions but not many of the problems they had, mostly integration. I found that about 70% of every implementation was taken up by integration, and I wanted to make a platform that was available to smaller companies out there. Even the products then were mainly targeted at the big tier 1 companies, and smaller companies that have the same problems didn’t really have a good solution. That was how it started, and then I found, and this was way back in 2003, that finding the people who could actually build it was incredibly hard. So, over the years, I morphed the idea, it grew, I had more epiphanies, and ultimately, a couple of years ago, I found the guys that I finally was convinced could build this thing, and they were convinced that this was a great idea to get involved with, and we started.
Kieran Chandler: And that brings us nicely onto our topic today, which is embracing uncertainty within supply chains. Demand is obviously the obvious example, but what other classes of uncertainty can we encounter?
Stefan de Kok: Well, everything in the future potentially is uncertain. So, if you’re in supply chain, you have to think about not just the quantities but maybe also the lead times, durations, quality or grade, yields, rates, pretty much anything that is going to happen in the future is uncertain to various degrees. So, we really have to look at the impact, not just of the average value of all these future things, but all the possible combinations of all the possible futures, which sounds very complex and it is, but it’s also what we need to do. And once something moves into the past, it’s almost certain. Right, even in the past, there’s some uncertainty. You’ve got data issues, and you might not even know if something truly happened a certain way, but for the most part, once it’s reached the past, it’s no longer uncertain to a large degree, and somewhere in the middle
Kieran Chandler: Joannes is going to be joining us as part of our discussion today, and Joannes, this idea of encountering an uncertain future is very much at the core of Lokad’s approach. So how do you approach these challenges? I mean, a traditional approach would be to use what-if scenarios, like optimistic and pessimistic scenarios.
Joannes Vermorel: The main issue with dealing with an uncertain future through scenarios is that it quickly becomes incredibly tedious and time-consuming. It takes so much effort to spell out those scenarios. Interestingly, with probabilistic forecasting, in a way, you’re kind of brute-forcing the problem. You might think that considering all the possible futures would be insanely difficult, but it turns out that if you have enough processing power, it’s actually a lot easier to implement the software and just run it compared to having a super complex system to manage many scenarios. It’s very interesting because not only is it mathematically elegant and concise for dealing with complex phenomena, but it’s also efficient for the supply chain, where you don’t have that many people. It’s relatively lean, both from the software development viewpoint and from the operational viewpoint, for the people who have to manage the system to actually operate a supply chain. That’s why I’m very interested and excited about this approach.
Kieran Chandler: Stefan, let’s look at some of the more traditional approaches people take. How are you seeing those classical approaches used to account for uncertainty?
Stefan de Kok: There are really two or three different ways that people deal with uncertainty. The first one is using buffers; the second is to respond to the uncertainty as it happens with a response mechanism, such as expediting; and the third one is just to ignore it. Everyone is doing some of each, and the question is how much should you spend in each. What generally happens is that with buffers, it’s all about the accuracy. If you get your buffer wrong, you need to overcompensate by responding, which is typically expensive. Finally, the parts that you cannot respond to are the ones you have to ignore, and those do the most long-term damage to a company. Customers get annoyed, you might lose market share, and it might even lead to lawsuits or bankruptcy in some cases if you ignore the customer long enough.
The most common buffers are lead times, capacities, and inventories. Companies will increase them because they know that if they’re short, they’ll run into issues that require a response. To give you an indication, many companies target 95 to 99 percent service levels, but when you measure their actual service, they’re achieving at best 90 percent, or usually, those targeting those levels are in the upper 80s. When you dig deeper, you find that even that number is usually driven by the response, not the buffer that they had initially planned. So they’re expediting at high costs and effort, and the amount of instability and firefighting is off the charts. Their inventory might only be providing them 73 percent service, even though they targeted 98 percent. This strains the company’s capabilities and erodes the margins, which I think is the status quo for most supply chains today, with a severe shifting of all the burden onto the response part.
Kieran Chandler: Let’s look a bit at the probabilistic approach, Joannes. It took a few years for you to develop it, so where did this idea come from?
Joannes Vermorel: Probabilistic forecasting for us was a journey. We actually started with classic forecasting, where you just forecast the mean. Then we had a client selling car parts, and we realized that if we were
Kieran Chandler: You know that it was so sparse, so intermittent, that basically forecasting zero just everywhere was actually, accuracy-wise, very, very good. It was obviously complete nonsense, and we came up first with quantile forecasts, which was like, “Oh no, you don’t want to forecast the average demand, you want to forecast something that has a bias on purpose.” And a forecast with an intentional bias is called a quantile forecast. That was the first step to say, okay, what should be just…
Joannes Vermorel: As Stefan described, you know, these situations with those C-items, you know, A, B, C or slow movers, how do you know if you want to have one, two units in stock or maybe three and not just have a min/max? So first, we realized that quantile forecast was a first step to even start to get results that were meaningful, you know, stop exit this situation where just forecasting zero is the best. It was making no sense whatsoever. And then we realized, “Oh, you want a forecast with a quantile, but what about tuning this quantile because you can tune how much bias do you want?” And then we went from quantile to quantile grids. Let’s have a series of biases incrementally increasing, and then we realized, “No, but we should probably have all the biases different.” And then we were basically down to quantile, contact grid, then probabilistic forecast. And then we, and by the way, there are separate literature, statistical literature, and it seems that many other people in the statistical community took the same journey as us, which is basically start with unbiased forecasts or predictions, move to biased ones, and then explore many biases and then try to do everything at once. And that’s a probabilistic forecast. Here you go.
Kieran Chandler: Okay, and Stefan, you also probably are one of the few people in the industry outside of Lokad who is also embracing this idea of a probabilistic forecast. So kind of what led to you having these ideas?
Stefan de Kok: Well, I guess this is where I talk about some of my epiphanies. The first one, which was that uncertain values cannot be represented by exact numbers, probably as early as 2006, I got that. But I hadn’t figured it out; it wasn’t a true epiphany yet. At the time, I hadn’t figured out how to make it work. It just doesn’t make sense to do it that way. And finally, I’ve been working on developing what I call a probabilistic arithmetic, and when I figured out how to make it work and I looked back, and I saw how something that looks so complex was actually solvable by something so elegant, it all fell into place. And that’s where I had my first aha moment. But at the time, I was keeping it a secret. I thought this is, you know, one of my key differentiators, so it certainly wasn’t pushing it yet.
It wasn’t until later, and this is one of those serendipitous moments in my career where I found myself needing money and looking for a job, and I found another company that was right in my hometown of Boston. One of the three companies in the world that do this, among Lokad, and at least at the time, and I found they’ve been doing this since the 1970s. And they have been proving this, but they have been keeping this a secret because this was their secret sauce. And I figured out a couple of things, and I think the key ones there are that there are many ways that you
Kieran Chandler: I think the key point is that there are many ways to achieve the same thing. Each of your approaches is very different, but the ultimate objective is the same. So, it wasn’t the way to do it that needed to be pushed, but the concept that it needed to be done. How do you feel about discussing and promoting this idea without giving away the secret recipe of what makes your companies special?
Stefan de Kok: I realized that I could talk about it, blog about it, and write articles about it without actually giving away the secret recipe of what makes us special. It’s important to make people aware that this is ultimately what all planning and forecasting will need to become in the next decade or so. One thing I’ve been pushing more recently is that traditional metrics are wrong as well. We need to change the metrics and use probabilistic forecasts in plants and probabilistic metrics to measure the value of that.
Joannes Vermorel: Absolutely, I agree with Stefan. When you have something probabilistic, you can simulate many possible futures, and you can challenge every single decision you make with its outcome as if you knew the future. It gives you a very elegant way to rank all your decisions and prioritize them. However, I believe the secret sauce, or the technology behind it, is less important. At Lokad, we’ve already discarded five generations of forecasting engines, each time believing it was the greatest thing of all time, just to realize two years later that there was a better way to do it.
Kieran Chandler: Joannes, it’s interesting that you mention the exploitability of future viability. Can you expand on that idea?
Joannes Vermorel: Sure. The fact that there is variability in the future can be exploited. It’s not just about protecting yourself and being more resilient; you can also take advantage of the fact that there is variability in the first place.
Kieran Chandler: Stefan, if we start to look at things from a customer’s perspective, how can we help them embrace this idea of accepting uncertainty? What strategies can they employ from a software perspective?
Stefan de Kok: The key part is that if you’ve got a complex engine and you get a black box output, it doesn’t lead to a lot of trust. The output should be self-explanatory. With probabilities, you can do an incredible amount. You can show that we don’t just think the answer is going to be a single number, but instead, we provide a range of possibilities that account for uncertainty.
Kieran Chandler: To be a hundred, we think the answer is going to be anywhere between a number of values, and there’s this distribution of how these values may occur. You can look at that at any kind of level, and it’s all about the visualization really of those results. I like to think of it almost like a car, you know? We are the mechanics, and the customer is the driver. In the old days, I knew how my car worked, and now it’s beautiful. I look in there, it’s beautiful, but I have no idea what makes it tick. Even the mechanic has to plug in a cable to connect it to his computer to figure out what’s going on. That’s how I think the solutions of the future and what we’re all bringing are all about: bringing that sophistication but making it easier for the user, for the driver to actually use it, and to get an output that they can use and that they can make decisions upon that are safer.
Joannes Vermorel: Okay, great. Sticking with that customer perspective, what does it mean for a company to change to that probabilistic approach? When you start to think about probabilities, it’s about trying to think about the sort of big forces that you’re trying to balance. What are the problems that you’re trying to mitigate? What are the bottlenecks that are going to hit you the most and hurt you? Usually, it’s very funny because those bursty forecast things just give you a way to quantify. It’s like the recipes that finally let you quantify what was frequently just obvious in terms of intuition. So, it’s not like it may produce fantastical insights. In my own experience, it’s kind of the opposite. It just puts stuff that was fairly obvious in the first place, but for the first time, the system gives you numbers that kind of match with the intuition in very mundane ways. For example, you have a product that is highly perishable; don’t put too much stock on it. You’re just taking massive risk with a super perishable product to have high stocks. If you take a classic forecasting approach, it’s just going to say, “Oh, just reach ninety-seven percent service level and be done with it,” and then you just create massive overstock with products as they expire. You take the probabilistic approach; the forecast might even be worse, actually. It might not even be super good, but it’s just more balanced, considering the risk that whenever you hit the stockouts, you know the expiration date, you have something that is very…
Kieran Chandler: Costly and thus, it steers the decision towards something that is much more sensible, which is not to overstock strawberries. So, I completely agree with the idea that the founder said that commercial detours have to kind of pursue simplicity. Although, to be fair, I don’t think that Lokad has had the most brilliant track record of delivering the simplest stuff ever, but at least they’re trying. Okay, Stefan, I’ll leave you with the last word. Based on what you’ve observed in the marketplace, would you say that people are ready to embrace the idea of accepting uncertainty, and what is it that really excites you for the future?
Stefan de Kok: I think we’re getting there. I’ve witnessed a lot of pushback over the years, and it’s been kind of an uphill battle, but I think we’re reaching the top. It’s becoming flatter, and I’m noticing less pushback. I think the market is realizing, and there’s a two-step approach. The first step is that an exact number is not the right way to go about dealing with uncertainty. Step two, that’s where I still see a little bit of friction. They think it’s overly complex. They say, “Everyone who says yeah, you could do it, but…” There’s always that “but,” and the “but” is often about big data. You need lots of data to do it probabilistically. Well, that’s just not true, right? You just need historical data, which you have in every ERP system, to solve the same problem that you would do deterministically.
The other concern that a lot of people have is how to deal with multiple possible futures. They think it’s going to explode the number of possibilities. However, one probabilistic forecast could lead to one probabilistic plan; you just need to express it in distribution. That is the part I think that people are still struggling with at this point. But I’m excited, I’m seeing the trend, and I’m seeing where it’s going. I’m seeing a lot more acceptance from some people that were vehemently against this whole approach earlier on. I see more and more people making the switch, having the epiphany, and so it’s a matter of getting to that critical mass, and it will be embraced by the mainstream. I’m sure I’m excited about that.
Kieran Chandler: Okay, brilliant. Well, we’re going to have to leave it there, but thank you both for your time. Thanks very much for tuning in, and we’ll see you again in the next episode. Bye for now.