00:00:08 Scenario analysis and comparison with probabilistic forecasting.
00:01:02 Scenario planning and its simplicity.
00:02:56 Limitations of scenario planning with software and handling uncertainty.
00:05:57 The alternative: probabilistic forecasting and its benefits.
00:06:48 Difficulties with implementing multiple scenarios in supply chain planning.
00:08:00 The limitations of scenario planning in supply chain management.
00:09:52 Probabilistic forecasting as an alternative to scenario planning.
00:10:38 Comparing probabilistic forecasting to high-resolution imaging.
00:13:22 Illusion of simplicity in scenario planning and decision-making challenges.
00:15:00 Example illustrating the inadequacy of averaging decisions across scenarios.
00:16:01 The limitations of traditional scenario planning.
00:17:27 Implementation challenges of probabilistic forecasting.
00:19:00 The benefits of scenario planning and its future.
00:20:58 Automating reconciliation between possible futures.
00:23:19 The importance of proper tools for probabilistic decision-making.
Kieran Chandler interviews Joannes Vermorel, founder of supply chain optimization software company Lokad, discussing the effectiveness of scenario analysis and its comparison to probabilistic forecasting. Vermorel believes that while scenario planning has limitations, such as being time-consuming and requiring significant input, probabilistic forecasting is a more efficient and effective approach to deal with uncertainty in the supply chain. He notes that exploring the unusual tools needed to operate over all futures at once is essential for adoption, and Lokad has engineered an algebra of random variables, a mathematical and statistical tool to deal with probabilistic futures. The recent unpredictability of supply chains highlights the need for better tools to deal with rogue scenarios.
In this interview, the host Kieran Chandler and Joannes Vermorel, founder of Lokad, a supply chain optimization software company, discuss the effectiveness of scenario analysis and its comparison to probabilistic forecasting. Scenario analysis was first developed by Shell in the 1970s and is a thought experiment where assumptions about the future are made and explored to see how things turn out, particularly in regards to supply chain matters. The purpose of scenario planning is to explore several alternative futures to make better decisions and create a more robust performance against unpredictable variations. Although supply chains are complex, scenario planning can simplify the process and can be supported by software tools. On the other hand, probabilistic forecasting is different in its approach and does not rely on assumptions. It deals with statistical probabilities and the uncertainty of the future. While both methods deal with uncertain futures, they differ in their approach, with scenario analysis being a more elegant and straightforward approach, and probabilistic forecasting relying more on statistical probabilities.
The discussion centers around scenario planning and forecasting in supply chain management.
Vermorel explains that scenario planning involves forecasting different scenarios based on various inputs and assumptions. These scenarios can help decision-makers plan for alternative futures and adjust their strategies accordingly. However, scenario planning also has limitations. For example, it can be time-consuming and requires significant manual input from experts.
Vermorel suggests that a more efficient and effective approach is to use probabilistic forecasting, which relies on computational methods rather than human expertise. This approach is particularly useful in situations where there are many variables to consider, and the number of possible scenarios is overwhelming.
Vermorel notes that while there are many consultants and vendors who advocate for scenario-based planning, it is rare to see these scenarios used in production. He suggests that this is because scenario planning requires a significant investment of time and resources, and it may not always be practical for companies to implement.
Overall, Vermorel emphasizes the importance of considering different scenarios and planning for uncertainty in supply chain management. While there are limitations to traditional scenario planning, probabilistic forecasting offers a more efficient and effective approach to this challenge.
The conversation focused on the role of scenarios in the supply chain. Vermorel believes that scenarios play a significant role in the supply chain team’s everyday tasks but are too costly. He believes that the core insight of dealing with uncertain futures is correct, but the classical supply chain perspective of doing a forecast is profoundly wrong. Vermorel argues that companies use scenarios because they lack better alternatives to deal with uncertain futures.
Probabilistic forecasting is a new alternative to deal with uncertain futures. Vermorel explains that probabilistic forecasting is different from scenarios because it takes a completely different perspective by leveraging large quantities of processing power that are cheaply available nowadays. The key idea behind probabilistic forecasting is to look at all possible futures, even if there is a very low probability of them happening. Vermorel believes that probabilistic forecasting is a better alternative to deal with uncertain futures than scenarios because it takes advantage of the large amounts of processing power available today.
Vermorel explained the concept of probabilistic forecasting, which is a mechanical way of putting probabilities on possible future events. With this, Vermorel believes that there is no longer a need for scenarios as all possible futures can be analyzed. He used the analogy of having a high-resolution image of the entire future rather than a low-resolution one with only a few pixels.
Chandler asked Vermorel about the differences in end-user perspective between scenario planning and probabilistic forecasting. Vermorel noted that scenario planning provides an illusion of simplicity but can be problematic when making decisions for multiple scenarios that can be inconsistent. Vermorel gave an example of a store selling books where the majority of clients are parents who only want one copy of a book, but occasionally, a school teacher enters and wants 30 copies of the same book.
Vermorel emphasized the importance of probabilistic forecasting as a tool for supply chain optimization and decision-making, as it provides a high-resolution image of all possible futures rather than just a few scenarios.
The discussion centers around the challenges of supply chain optimization. Vermorel is the founder of Lokad, a software company that specializes in this area. The conversation begins with a discussion about the appropriate amount of stock to keep in a store for a given book. Vermorel argues that averaging the amount of stock needed does not make sense and that the traditional approach of pouring more manpower into the problem exacerbates the issue. He explains that scenario planning offers a potential solution by allowing for the management of multiple futures, but it requires more than just copying and pasting logic. Implementing a probabilistic forecasting approach is much more challenging, but for supply chain software vendors like Lokad, it offers a significant opportunity to improve their systems. Vermorel concludes by noting that reconciling the results that emerge from scenario planning is a critical challenge that needs to be addressed.
They discuss scenario analysis and its limitations in supply chain management. Vermorel explains that companies often don’t realize they have a problem until they implement a system and run into edge cases that make it non-viable. Despite this, vendors are incentivized to sell features that look cool and work during demos, and as long as people can’t imagine an alternative, scenario analysis seems like the best option. Vermorel believes that the probabilistic forecast, which assigns probabilities to scenarios, is a natural extension of scenario analysis, but the real problem lies in completely automating the reconciliation between all possible futures. To solve this problem, Lokad has engineered an algebra of random variables, a specific mathematical and statistical tool to deal with probabilistic futures. Vermorel emphasizes that exploring the unusual tools needed to operate over all futures at once is essential for adoption, as it is not enough to realize that it’s possible to look at all possible futures with probabilities. Finally, they discuss recent events that have shown the unpredictability of supply chains and the need for better tools to deal with rogue scenarios.
Kieran Chandler: Today on Lokad TV, we’re going to discuss its effectiveness and whether it can be replaced by alternative methods such as probabilistic forecasting.
Joannes Vermorel: So, these two methods are similar in the sense that you want to deal with uncertain futures. Because you don’t know for sure what the future will be, you want to explore options. However, the way you do this exploration is very different.
Kieran Chandler: Today, we’re going to be looking at scenario planning in more detail. Perhaps a good place to start is by explaining how it works and what it is.
Joannes Vermorel: Scenario planning is conceptually very simple. You’re just doing a what-if thought experiment, saying the future is going to be like this. Let’s assume that the future is exactly like this, and based on this initial assumption, let’s see how things turn out in terms of supply chain perspectives if this future happens to be the one. In a way, it’s very elegant and simple, and you can repeat the exercise. The idea behind scenario planning is that instead of just having one true forecast, we are going to explore several alternatives. By exploring several alternatives, you can make your decisions and performance more robust against variations that you cannot predict, but expect.
Kieran Chandler: It’s interesting because we know that in supply chain, demand might be one variation, but there also might be other related variations, like lead time. Can this method combine those multiple variations?
Joannes Vermorel: The key idea of scenario planning is that it makes things very simple by mentally sticking to something that is very similar to the past. It is literally saying the demand will be numerically this and that. Due to this simplicity, it turns out that from a software viewpoint, you need tools to support these lines of thought, especially when dealing with the complexity of thousands of products, tons of locations, etc. The interesting thing is that, from a software perspective, scenario planning is like a cut-and-paste exercise with your setup and logic dealing with the classic forecast.
Kieran Chandler: Can you explain what scenario analysis is and how it’s used in supply chain optimization?
Joannes Vermorel: Sure, scenario analysis is based on forecasting numbers that represent possible future scenarios. Based on those forecasts, you can infer decisions. If you want to have an alternative scenario, it’s just a different forecast with maybe a bias. For example, a downward trend. You can then infer all the decisions and compare them to the first set of decisions you made for the median scenario, which represents your usual classic forecast.
The fact that it’s very easy to implement this sort of process with software where you can just multiply scenarios doesn’t necessarily mean that it will do everything you want. In particular, you depend on experts and manual tweaks to adjust the scenarios. You tend to adjust the scenarios according to simplistic variables such as inflating or deflating future demand. But there are plenty of other areas that are also uncertain, such as lead times.
The problem with tweaking scenarios is that you have tons of combinations and it becomes overwhelming. Fundamentally, the choice of scenarios is a very human-driven affair. Humans are supposed to cherry-pick those scenarios.
The alternative approach is to take a more computational approach, and that’s where probabilistic forecasting fits in.
Kieran Chandler: So, is that why it’s better to leave it up to a machine rather than a human being?
Joannes Vermorel: Yes, scenario analysis is an interesting way to tackle the fact that you don’t know the future perfectly, but it suffers from the problem that you do cherry-pick a few scenarios. The reality is that the classic approach for supply chain forecasting and planning is already very time-consuming, even for just the primary scenario. For large companies, it already takes entire teams of planners and forecasters to get the job done. If you want to add more scenarios, you face a situation where you need a nearly linear increase in manpower to fuel the process. That’s why, in my experience, although there are tons of consultants and vendors who advocate for scenario-based planning, I have very rarely seen those scenarios truly used in production. It’s more like the exception rather than the norm. If we look at daily operations, it’s actually quite rare that they are used.
Kieran Chandler: To see companies where scenarios play a significant role in the everyday task of the supply chain team is rare, and I believe that the reason for that is that it’s just way too costly. But this is a technique that’s been around since the 1970s. So why is it that consultants and vendors are still pushing it, and why is it something that companies are evidently still using?
Joannes Vermorel: First, I believe that the core insight, which is the need to deal with uncertain futures, is profoundly correct. So, indeed, the classical supply chain perspective, where you just do a forecast and say “this is it, this is the future” like next week we will be selling 155 units of this product, is fundamentally flawed. Intuitively, you see that there is something profoundly wrong with this approach; you need to deal with the fact that the future isn’t known. Scenarios are an answer, albeit a poor one, to this problem, and lacking better alternatives, it is very tempting to use them. There is this motto that when all you have in your hand is a hammer, everything else is a nail. So, if you don’t have any better tool, you will use what you have, which is a hammer, even if it’s actually a screw that you’re trying to deal with.
Now, alternatives have emerged that rely on things that were non-existent four or five decades ago, like large quantities of cheap processing power. That’s the essence of probabilistic forecasting. It’s basically taking a completely different perspective at the problem by leveraging the large quantities of processing power available nowadays.
Kieran Chandler: So why is probabilistic forecasting maybe so different? Because you’re still kind of looking at an alternative future. I guess with probabilistic forecasting you’re assigning a probability of it maybe occurring, but on the surface, they don’t seem to be actually quite that different because you’re just looking at possible alternatives.
Joannes Vermorel: The key idea about probabilistic forecasting is that you will have machine-generated futures. You will look at all possible futures, or at least all the numerically relevant futures. Even if you have immense computing resources at your disposal, there is no point in assessing the consequence of something that has only a chance out of a trillion to happen. It’s just not a good investment of the processing power that you have. But if we only consider things that have at least a chance of happening, let’s say at least one chance out of a million every year to happen, it’s still very improbable, but it’s not vanishingly improbable. Then there is plenty of raw processing power to deal with those sorts of events through modern computers.
Kieran Chandler: Decades from now, with just the mundane sort of computers that you have at home or even in your smartphone, probabilistic forecasting is literally just a mechanical way to put probabilities on tons of possible future events. As soon as you have that, you realize that you don’t need scenarios anymore. You can replace all those scenarios that were cherry-picked among all possible futures with something that brutally analyzes all possible futures. It’s like transitioning from a video camera with only four pixels in your image, to a complete high-resolution image of the entire future where you see all the pixels.
Joannes Vermorel: What difference does it make? Just ask yourself, if you have a landscape and you have only five pixels, it’s very hard to infer what it looks like. If you put in a lot of effort, you might have 20 pixels on your image, and that’s going to be very tedious. But when you transition from 20 pixels to 4 million, you go from something that barely makes any sense to a high-quality picture that completely makes sense. In practice, even if it’s a gradual process in theory, it’s completely different.
Kieran Chandler: Let’s talk about things from the end user’s perspective. With scenario planning, maybe with just your four pixels, you have an idea about what’s happening at the beginning and the end, and it’s very logical. But this idea of looking at a probabilistic forecast and all possible futures is much less logical and more difficult to understand. Is it easier for the end user?
Joannes Vermorel: With scenario planning, you have an illusion of simplicity. You pick one scenario, and you have one set of decisions that make sense for that particular scenario. It’s relatively straightforward. But the problem is that you look at another scenario and get another set of decisions. If you look at a third scenario, you get yet another set of decisions. What are you going to do with all those decisions that can be and are, in practice, completely inconsistent? The naive answer would be to make some kind of average, but why would the average be relevant?
Just to give you an example, imagine that you have a store selling books for schools. The majority of your clients are parents who have one child and just want one copy of the book. In the same store, once in a while, a school teacher enters and wants 30 copies of the book.
Kieran Chandler: The question is, does it make sense to have a stock of five copies of a book in a store?
Joannes Vermorel: If you only want to serve the parents that once in a while come to the store, then yes, five copies might be enough. However, if you want to serve the teachers, then you need probably 35 copies of the book. You will need 30 to serve the teachers and then five more to serve the parents. But if you average and say, “Oh, 25 should be enough,” no, 25 is kind of bad because it’s way more than what you need for the parents and it’s still not sufficient for the teacher. So, you see, averaging usually doesn’t really make sense. And if you have non-linearities like MOQs, minimal order quantity, it doesn’t work out. So basically, you have an appearance of simplicity when you look at those scenarios where you pick one scenario and say, “Well, there is a simple solution here.” You pick another scenario, and you have another simple solution. But then you end up with a problem of how do you reconcile all those scenarios. The reality is, well, the traditional approach is just to pour even more manpower to the problem. So, you see, you already had something that was super intensive in terms of manpower. Every single scenario, you need more planners, more forecasters. But then you’re creating another problem, which is, you know, now you need another team to just reconcile all the stuff. So, you see, it’s even worse than having a linear increase in the amount of staff you have. It’s like a hyper-linear increase in the amount of stuff that you need to deal with a great number of scenarios. And that’s very, very bad.
Kieran Chandler: I think maybe one of the benefits of scenario planning is just a case of copying pasting a bit of logic and from an implementation perspective, it’s just about throwing more resources at it. Does that mean that the implementation of a probabilistic forecasting approach is much more challenging, and that’s why companies aren’t so interested in it?
Joannes Vermorel: I mean, for as a software vendor, absolutely yes. Literally, imagine you’re a software supply chain software vendor, and what you do is you’ve engineered some kind of supply chain forecasting and planning system. It means that at the very least, from the classical perspective, your system is able to manage a classic forecast. So, it’s about to assign a daily, weekly, or monthly quantity for X period ahead for any single skew or product. This is what a classical forecasting product looks like. You have everything in place to manage one future and say, “This is it. This is how things will happen.” Now, if you want to transition towards scenario planning, you only have to basically say, “Well, I’m just going to manage a second instance of this future, and I’m just going to refer to that as a second scenario.” In terms of logic, in terms of code, it’s literally the exact same code where you just need to add one extra dimension, which is the scenario dimension, and then this is it. You already have your scenario planning system in place. But there is a catch. The catch is that you’ve not done anything to reconcile the results that emerge from those.
Kieran Chandler: Scenarios, but the reality is that companies are usually not going to realize, you know, that they actually face this problem until they implement the system and they run into all the sort of edge cases where you cannot average the results given by various scenarios, where it’s actually nonsensical to average the results.
Joannes Vermorel: If you do like three examples and the vendor will be quite good, you know, I mean first, the interest of the vendor is very cheap to implement, so why not implement a feature that is cheap and looks cool during the demos? Plus, when you just do a very simple pilot, chances are that you will not run into the edge cases that I’ve just described, and thus averaging the scenario will likely work for you. But as soon as you grow into a more complex situation, you will realize that you have endless edge cases that emerge, and thus that makes the whole thing non-viable. But at this point, the vendor is already happy because the vendor has already been paid. You know, there are all the wrong incentives in place on how you pay your supply chain software vendor, which we discussed in the previous episode. Plus, again, as long as people cannot imagine an alternative, it still seems like the best thing that there is for a lack of anything better.
Kieran Chandler: Okay, so let’s sort of start bringing things together a bit then. You say that people can’t imagine an alternative, and now we’re kind of in a position where there is an alternative with the computational power that we have. There is an alternative method, so can you see something such as scenario analysis kind of dying out at some point?
Joannes Vermorel: Yes, I mean, although frankly, there is still a long way ahead for Lokad to evangelize quite literally the rest of the world on that. The problem is not the probabilistic forecast per se because the idea that you’re going to have many scenarios with probabilities is relatively straightforward, and it feels like a very natural extension of the scenario approach, just with many scenarios, just like adding pixels to an image, where the quantity would be the probability.
Kieran Chandler: So instead of the color, you could think of this as my scenario, and there’s a probability assigned to this scenario. But the problem is that once you have all those possible futures, you need to think of ways to completely automate the reconciliation between all those possible futures. And that’s where things start to become really tricky and bizarre from the Lokad perspective.
Joannes Vermorel: You know, the solution that we have engineered is called an algebra of random variables. So, you need some kind of very specific mathematical statistical tooling to deal with those sorts of concepts, not just on the forecasting side, but on the decision-making side. Actually, if you look more at the Lokad technological stack, you would see that forecasting is just one relatively small piece of the picture nowadays. The bulk of the complexity, techniques, tools, practices, algorithms, and the rest lies in how do you actually optimize your decisions while facing all those uncertain, probabilistic futures.
And that’s the problem, I believe, for further fostering adoption. People have not only to realize that it’s possible to look at all those possible futures with probabilities, but they also have to explore things that are, I would say, very unusual - the sort of tools that you need to operate over all futures at once. Because without those tools, you are basically back to square one, where it takes an enormous amount of time for people to sort out all those possibilities. The trick is that, at Lokad, we don’t sort out all those possibilities; we maintain all those possibilities during the entire process. So, it’s not exactly easy, but it’s actually quite straightforward if you have the proper tooling.
Kieran Chandler: Okay, we’ll have to wrap it up there. But I guess, based on recent events, we kind of know there are so many rogue scenarios that can occur that humans probably could never predict. So that’s everything for this week. Thanks very much for tuning in, and we’ll see you again in the next episode. Thanks for watching.