00:00:00 Introduction and definition of terms
00:02:40 Uncertainty and cost of safeguarding in supply chain
00:03:54 Risk management and minimizing waste
00:05:30 Irreducible risk and opportunities in supply chain
00:07:37 Supply chain vs manufacturing perfection
00:09:35 Risks and opportunities in supply chain and competitors
00:14:09 Problems with static approach in supply chain
00:15:56 Predictable mistakes as business practice
00:18:46 Engineering agility in supply chain
00:21:20 Dollar value of risk and opportunities
00:23:36 Financial optimization of supply chain risks
00:26:37 Lokad’s approach to probabilistic forecasting
00:29:53 Risk of massive disruption and regional disasters
00:31:59 Factoring risks into daily supply chain decisions
00:34:08 Risk of losing big clients and correlation of risks
00:37:03 Distortion in map projections and mathematical models
00:42:31 Building forecasts and time series risk
00:45:20 Stochastic optimization and probabilistic approach
00:48:36 Decomposing economic drivers for supply chain decisions
00:51:44 Companies often surprised by past events
00:57:00 Damaging forecasts and cross-entropy in forecasting
01:00:00 Importance of actionable risk assessments
01:06:49 Financial risks of inventory distribution
01:13:54 Cost of promotions and IT dependencies as a risk
01:17:22 Difficulty of modeling customer psychology
01:24:26 Evaluating time series forecasts
01:27:33 Risks in mainstream supply chain software
01:29:30 Closing thoughts and call to action
Conor Doherty, LokadTV host, and Joannes Vermorel, founder of Lokad, discuss the inherent risks in supply chain management. Vermorel emphasizes that the primary risk is the uncertainty of the future, which is irreducible and beyond control. He notes that every decision involves a trade-off between risk and reward, and that zero risk is unattainable. Vermorel also highlights the opportunities that can arise from these risks, such as capitalizing on market shortages. He advocates an agile and opportunistic mindset, and the use of probabilistic forecasting to mitigate risk. Vermorel and Doherty conclude by agreeing that even small companies can benefit from risk management, leading to increased margins and cash flow.
In a conversation between Conor Doherty, the host, and Joannes Vermorel, the founder of Lokad, a software company specializing in supply chain optimization, the topic of risk management in supply chains is explored in depth. Vermorel explains that the primary source of risk in supply chains is the uncertainty of the future. Unlike in manufacturing where risks can be eliminated by perfecting the process, supply chain risks are contingent on future conditions that are unknown and irreducible.
Vermorel emphasizes that every decision in supply chain management involves a trade-off between risk and reward, and the uncertainty about the future is irreducible. He also notes that zero risk doesn’t exist in supply chains. Even with a perfect manufacturing process, there’s always a double-digit percentage risk that a product will not be sold in the market five years from now.
Vermorel reiterates that the source of risk in supply chains is the uncertainty of the future. He explains that this risk is irreducible and unlike other areas like accounting where risks can be eliminated, supply chain risks are beyond control and can only be mitigated. He also points out that while there are risks in supply chains, there are also opportunities. For instance, having a lot of stock when the market is facing a shortage can lead to a nice profit.
Vermorel agrees with Doherty’s observation that the risks in supply chains seem to be more common and have predictable financial effects. He emphasizes that supply chain management involves dealing with a lot of risks and opportunities that are mostly beyond control, and the only thing that can be done is to make decisions that balance these risks and opportunities.
Vermorel explains that both risks and opportunities can arise in supply chains. For instance, a competitor entering or exiting the market can either lower or raise prices, affecting profitability. He also gives an example of a European e-commerce company that capitalized on a surprising event to make an exceptional profit.
Vermorel explains that in supply chains, it’s important to have an agile and opportunistic mindset. He notes that while it’s possible to establish practices that take advantage of predictable mistakes, it’s also important to be prepared for emerging risks.
Vermorel discusses the cost of defects in car manufacturing, noting that in supply chain management, there are no hard constraints and everything is eligible for financial optimization. He explains that forecasting is used in supply chain management to mitigate risk. However, unlike defects in car manufacturing, forecast inaccuracies will never be completely eliminated.
Vermorel introduces the concept of probabilistic forecasting, where every possible future has a probability. He explains that risks like a 2% chance of losing clients can be factored into probabilistic forecasts by forecasting the behavior of the clients, rather than the demand for the products.
Vermorel argues that the quality of a risk-driven approach should be judged by the end results, not by the quality of a probabilistic forecast. He emphasizes that the main concern should be how much money was made or lost, not the accuracy of the forecast.
Vermorel suggests that cross entropy, a measure of probabilistic forecast accuracy, is as abstract as percentages, but has good properties for learning high-quality models. He emphasizes that the end game results, which result in reduced errors in terms of money, are the most important metric.
Vermorel suggests that such risk management divisions are often bureaucratic and their assessments have no consequences if they don’t adjust every single decision in the supply chain. He argues that if risk assessments don’t impact purchase decisions, they are being ignored.
Vermorel agrees, arguing that while both are important, macro decisions are often a gamble, while daily decisions can be quantitatively assessed.
Vermorel discusses the classic perspective of supply chain management, where each store is treated independently and a safety stock is maintained to cover potential shortages. However, this approach can lead to issues such as running out of stock at the warehouse level. He criticizes the traditional approach of allocating safety stock to stores sequentially, which can result in some stores being fully stocked while others receive nothing. This can lead to unserved demand and excess stock in certain stores.
Vermorel suggests a better approach would be to spread out the inventory so that all stores run out of stock at the same time, maximizing sales. He proposes a probabilistic approach that takes into account the network and interdependencies between all stores and the available inventory.
Vermorel discusses intangible risks, such as customer expectations and brand value. For example, offering discounts can lead to customers expecting future discounts, which can be difficult to quantify and manage. He also mentions other classes of risk, such as IT dependencies, which can impact the supply chain.
Vermorel explains that promotions can lead to customers waiting for future promotions before making purchases, which can be difficult to model and manage due to the long-term nature of customer behavior. He suggests that a risk-driven approach is more compatible with these types of guesstimates compared to traditional time series planning forecasts.
Vermorel challenges the notion that only large companies can afford to manage risk, arguing that ignoring risk can be more costly. He suggests that probabilistic forecasts can be more effective and easier to deploy than traditional time series forecasts. Vermorel argues that even small companies can benefit from risk management, as it can lead to increased margins and cash flow.
In conclusion, Vermorel agrees with Doherty’s summary, arguing that many companies face preventable catastrophes due to ignoring risk. He suggests that a better match between reality and supply chain management can lead to a higher degree of automation and fewer people needed to manage the process.
Conor Doherty: Welcome back to LokadTV! Risk is systemic in supply chain. From a certain perspective, every single supply chain decision presents potential classes of risk, either directly or indirectly. Here to explain why - and importantly how to avoid them - we have Joannes Vermorel, founder of Lokad.
So Joannes, to quote the Great American thinker George Costanza, in order to manage risk, we must first understand it. And to understand it, we must first define it. So in the context of supply chain, what exactly is risk management and how does it differ from risk management in other fields like manufacturing?
Joannes Vermorel: The primary source of risk in supply chain is the future that you don’t know. In manufacturing, it’s mostly about having the correct process. If you have the correct process, you can possibly produce a billion parts without ever facing any defects. So potentially, manufacturing-wise, you can eliminate the risk. The risk is not so much contingent to the future. For example, if you have a faulty process that creates a fire hazard for your factory, it’s just a matter of time before the factory catches fire.
In supply chain, it’s literally the future you don’t know and you can’t really safeguard all the possibilities because that’s too costly. There is always a possibility that the demand for specific products spikes by a factor of 20, but should you have in stock 20 times more than what you expect to sell, considering the applicable lead time and whatnot, just because there is this one remote possibility of facing this super unpredictable spike of demand? No.
Fundamentally, I would say, unlike some other domains, in supply chain, the sources of risk are the future conditions that you don’t know and every decision that you take is kind of a trade-off about the sort of risk and rewards also that it comes together about this uncertainty about the future. And I would say the uncertainty about the future is irreducible, unlike let’s say the uncertainty such as the physical assets in your manufacturing process in a factory.
Zero risk never exists. Although if you’re looking at a manufacturing process, you can get very, very close to zero. I mean it’s not absolutely zero, but it’s very, very close. Supply chain-wise, look at any product, there is always a double-digit percentage of risks that this product will not be sold anymore in the market five years from now. There are very few products that you can say with absolute confidence that they will still exist five years from now, especially if we take into account that the product can be replaced by a variant, which still counts as a different SKU.
Conor Doherty: So, to summarize, are you saying that risk management in supply chain is purely a matter of minimizing the wasted resources or is it just purely a financial concern?
Joannes Vermorel: The source of risk is literally the fact that you don’t know the future. If you had a magic crystal ball that would tell you the future, you could in theory have almost a risk-free supply chain practice, assuming one has enough money.
This source of risk is irreducible and feels strange in many other areas. For example, in accounting, you have a risk of having accounting mistakes, but that’s what those accounting practices are for, to essentially eliminate this risk. When you think in terms of risk associated with incorrect accounting practices, you really want to make that super rare.
In supply chain, you don’t have the option. No matter if you’re good, if you have the correct practices, the risk is irreducible. You might have a war, lockdowns, fires, all sorts of events that are just beyond your control that will massively steer the demand one way or another. That’s the primary source of risk, that you don’t know and everything that you can do is about mitigating those risks. But also, as there is risk, there are also opportunities which do not exist in other areas like accounting practices.
For example, if you happen to have a lot of stock of something while the market is facing a shortage, you can make potentially a nice profit in selling this inventory at a premium.
Conor Doherty: It sounds like a lot of those issues in supply chain will be much more common than the example you gave of hiring an incompetent or morally wrong person. They’re vanishingly rare, but presumably the classes of risk you’re talking about in the context of supply chain, like lead times being extended or even hastened by a couple of days, those are presumably quite regular and have predictable financial knock-on effects.
Joannes Vermorel: Yes, and they don’t depend on you. That’s also one thing that is very different. If you’re in a manufacturing process in a plant and you have defects, fixing the process so that you don’t have defects anymore is all on you. You can potentially go to those zero defect states which is perfection, or you can get very close to perfection.
Again, if we go to the world of supply chain, not really. I mean those things, by definition, if you have a lead time, you have a supplier and this company is beyond your control. And even if you internalize, you may have a transporter and it’s still beyond your control. And even if you internalize the transporter, the road might be cut because the highway is flooded or there is something else and again it’s beyond your control.
So, what makes the supply chain practice so specific is that you’re dealing with a lot of risk and conversely a lot of opportunities and they are mostly beyond your control. So the only thing that you can do is to make decisions that properly balance those risks and opportunities.
Conor Doherty: So, when you talk about opportunity in the context of risk, you mean missing opportunities?
Joannes Vermorel: Yes, a competitor can suddenly enter the market and lower the prices. That’s a risk. So, you may be forced to lower your price in turn and then you’re less profitable than you would expect or even maybe you’re not even profitable anymore. But just the opposite can happen. A competitor may exit your market. In this case, well, you can raise your price and you’re more profitable than you would expect.
Every time you think that there is a risk, there is the opportunity. If there is a flood, maybe your warehouse is going to get flooded or maybe one of your competitor’s gets flooded. So, when people think about risk again in manufacturing settings, you have a clear goal which is this perfection. So when you think risk, it’s necessarily there is no specific really opportunistic random upsides.
But in supply chain, this may happen. You may have thousands of products and for some random reasons, competitors just make mistakes. They don’t have the proper amount of stock, they don’t have the right capacity or they have the wrong allocation and then there are opportunities.
For example, a large European e-commerce company, one of their techniques was that they would start selling fashion products and they would identify very quickly in the season the bestsellers, literally within a day or two. And what they would do is they would immediately pass a gigantic order to the original brand and they would corner all the stock.
It was surprisingly well received, and thus they were saying, “Okay, if this amount of sales comes as a surprise to us, most likely it comes as a surprise to the original brand. So, what will happen if we place a massive order? We will be sitting on a large pile of inventory while everybody else is running out of stock. We can sell the same products at a slightly higher price than the normal price, and we will still sell everything without incurring any penalty related to end-of-season sales.”
So, you see, the idea is that there is a surprising event, a product sells more than experts would expect, and then, if you’re smart, you can turn that into an opportunity to corner the amount of inventory that happens to be still available and then make an exceptional profit on this product. So, you see, the risk is there, but there is also the opportunity that arises.
Conor Doherty: Got it, thank you. That example is quite interesting because it opens up a potential fork in the conversation. If I understood correctly, the example you gave was one of a reactive response to opportunity. This clothes fashion seller spotted an opportunity and very agilely responded to this opportunity. So that was a reactive approach to managing opportunity and avoiding risks. Is that the best you can do in supply chain, or is there a proactive mechanism for anticipating these kinds of events?
Joannes Vermorel: I would have a twofold answer to that. First, you’re very correct. It came with an agile mindset, an opportunistic mindset, and it applies equally for risk and opportunities. The interesting thing is that if you come from a perspective like manufacturing, this is not the perspective that you adopt. You just want to eliminate risk. It’s a static problem. Either your process has no risk, no defect, no hazards, and you’re good, or it doesn’t, and you have to fix it.
Here in supply chain, the interesting thing is that when you try to approach risk with this sort of stationary mindset, you think of it as something that you could fix once and for all. But the problem is that it doesn’t work because if you have something that is completely static, then you can’t capture the opportunities anymore. But the reality is that you can’t react to the emerging risk as well. It’s completely symmetric. So you have opportunities that arise, but there will be risks that just bubble up and surprise everybody, and you need to react quickly as well. So it’s symmetric.
Now, what does it mean to be prepared? As I was mentioning this example on this large European e-commerce company that plays this game of cornering brand inventory, this is an established practice. They know that due to the fact that a large fashion brand is going to have a collection of maybe 20,000 distinct variants, mistakes will be made. It’s a certainty. You don’t know which one, but the idea that a sizable brand could get everything properly sized in terms of inventory is a relatively safe bet to say that mistakes will be made. And thus, you can establish a practice where you turn those predictable mistakes into your advantage.
Conor Doherty: When you talk about engineering a process in a company like someone dealing with fast-moving consumer goods, how exactly do you instantiate that? Is it a top-down or a bottom-up sort of process? I mean, taking advantage of those opportunities, how do you instantiate that sort of process?
Joannes Vermorel: Like most things in supply chain, it has to be top-down to some extent. You can’t expect people at the very bottom to have any way to re-engineer the organization itself. For example, if you decide that your process is like SNOP (Sales and Operations Planning) and then you have quarterly sessions for SNOP where you spend two months to establish your new forecast and have everybody agree on that, establish a big consensus, and then everybody is surveyed and then you need to compile all the results and then you need to re-translate the forecasts which are per week per category into something that makes sense in terms of decisions, you’re in a situation where it doesn’t matter if people at the very bottom are agile or not. The process and the organization itself prevent any kind of agility anyway. So, to a sizable extent, if you want to be agile, this has to be engineered from the top so that this agility may even happen. But then, once you have engineered something where it becomes a possibility, then yes, it’s a much more bottom-up thing because then it’s about whether the various teams take advantage of this newfound agility.
Conor Doherty: It also occurs to me that there’s another way to approach the idea of risk and opportunity. If you just invert the example you gave, instead of focusing on the company who were selling these t-shirts that were going like hot cakes and decided to corner the market on this, that’s them taking advantage of an opportunity. From the perspective of the supplier, if you’re in a situation like that where all of a sudden out of nowhere Joannes’ clothing store calls me and says, “Oh, we want all the t-shirts that you have, all the black t-shirts, we’ll buy them right now,” is that something you should be wary of? Because again, there is risk and opportunity there. As a supplier, should I sell? It’s a guaranteed sale right now, today, I clear everything. Or should I investigate why he would be trying to buy these right now? Is there something else at play here?
Joannes Vermorel: It really depends on whether you can even afford to spend time investigating. If there is an EDI connection and purchases are completely automated and there is not even somebody in the loop, it really depends. But having adversarial behaviors all over the place is just another day in supply chain. Your suppliers are your best partners and potentially your competitors because they can also limit your profit. They can also become in time competitors, decide to have their own brands, etc. And the converse is true. If you’re a brand, you can decide to internalize and suddenly you compete with what was your former suppliers. So, there are no general rules, it really depends. But the interesting thing is that in supply chain, you can put dollars or Euros of rewards and opportunities on those sorts of things.
Again, if you go back to car manufacturing, how much does it cost to have a defect that kills one person? The answer is way too much. So, you see, it’s not the sort of thing where you are going to do fancy engineering because it’s mostly not acceptable. So yes, in theory, economists would tell you the cost of a human life in the US according to various things is let’s say five million, whatever, you may even make a case for that. But the reality is that nobody is going to do really serious engineering. They would just do whatever they can so that those problems where you have somebody dying just do not happen. And so, there is no real financial engineering because again, if you take it from the manufacturing angle, you just want to avoid by design those sorts of problems and you’re not trying to optimize your risk in the sense of balancing the pros and the cons, you just want to eliminate that. But in supply chain, you can’t, and it’s going to be a real trade-off. Whatever you do has a cost, there is a reward, and it’s only shades of gray. So, it’s not so. You can always have a little bit more stock, you can always operate with a little less stock, and you can even try to operate with zero stock at all and you just do back orders all the way. So, it is, you have a lot more flexibility and also you have very little hard constraints in supply chain. As long as you’re willing to pay, there are almost no constraints whatsoever. You want more storage space? If you’re willing to pay for it, you can actually pay to have a second warehouse being built. So, ultimately, all the sorts of constraints, all the sorts of risk and rewards, they are kind of soft and thus they are very eligible to financial optimization as opposed to life and death situations where people would say, “No, we are not going to do a financial optimization of that. It has to be a categorical answer. We just don’t want that.” So, supply chain has this luxury that the vast majority of the problems are actually soft problems where you can go from super bad service to super good service and the whole spectrum is possible and the cost structure evolves as you go for better quality of service or worse quality of service.
Conor Doherty: The example you gave of car manufacturing actually presents a very nice segue because I know in car manufacturing, for example, Ford, they manage risk, particularly with their autonomous vehicles, by using digital twins. They build a digital version and a digital environment and then, using algorithms, they subject the theoretical autonomous vehicle to a battery of tests and assess their risk without ever having to produce a prototype in the real world. That’s one step for managing their risk. Is there anything like that for supply chain? Because again, it’s not a physical product in and of itself, though it’s composed of many moving parts.
Joannes Vermorel: That’s the interesting thing. That’s what you try to do in a way with forecasting. You try to mitigate this risk that you have about this uncertain future through forecasting. Ideally, if your forecasts were perfect, you would just eliminate this risk. That’s why there are many supply chain practices that treat forecasting accuracy as car manufacturers treat defects in their brake pads, as something that you should eliminate.
But the problem is that unlike defects in brake pads, where you can potentially bring this rate of defect to one per billion, so that it’s so low that it’s inconsequential, the forecast inaccuracy is never going to go to 0.01 of error. It’s usually going to be stuck, if you look at the sort of granularity that makes sense forecast-wise for the decisions, so basically per SKU per day, you’re going to be stuck with widely inaccurate forecasts, like 50% inaccurate on average, per day per SKU, if you look a few months ahead.
The interesting thing is, what do you have in terms of tooling and processes and methodologies to deal with these classes of risk? That’s pretty much what Lokad is doing with probabilistic forecasting, just for this reason. That’s a way to embrace this uncertainty. But that is very different from the classical paradigm that just assumes that the forecast will be accurate and where if there are inaccuracies, this is treated as a defect that should be resolved.
The Lokad approach, probabilistic forecasting, is that we don’t assume or even expect that those inaccuracies will ever go away. What we have is probabilities. We may improve our models to have probabilities that are a little bit more concentrated, so we have a vision that is a little bit sharper about the future. But the overall perspective is that it’s going to remain extremely fuzzy and extremely uncertain, no matter what.
Conor Doherty: I do want to plant a flag there because I think there’s an important point and I just want to amplify it. When you talk about forecasting future demand, most people would hear that as just looking at previous sales data and coming up with a number, like on a time series. Is your position that the probabilistic forecasting approach will factor not only the historical data but the other classes of risk we’re talking about, like extended lead times, a boat gets stuck in a canal, or something like that, and merges them together?
Joannes Vermorel: Yes, absolutely. That’s why we at Lokad typically speak of predictive modeling rather than forecasting. Forecasting, in theory, you could forecast anything, but the reality is that when you say forecast, the default expectation is that you’re talking about the demand or the sales. That’s 99% of the situation when people say we have a forecast, they mean a forecast of the sales or the demand. But the reality is that anything that is uncertain about the future can be anticipated and thus we have this predictive modeling.
The interesting thing is that there are plenty of things where you can model risk even if you don’t have really data. For example, war in Europe. If you look at the last 100 years, there has been like one major war every half a century. So, if you look at it, that means that there is every year there is like a two percent chance that there will be a war that would impact you. You can go back five centuries in the history of Europe and that’s something that has been happening over and over.
I hope that the risk of actual war for Western Europe is fairly low at the moment, but again, if you take an historical perspective, saying that there is a two percent chance to have a massive disruption is not relatively insane. Look at what is happening in Ukraine. The risk is definitely real and 20 years ago it was in ex-Yugoslavia. So those sorts of things do happen and you don’t need to have precise data to say well we can put a two percent risk of a major disruptive event.
You might be dependent on the region, you might be flooded, you might have fires. There are plenty of risks where you can do a ballpark assessment. It’s better to do that as opposed to pretend that those risks don’t exist at all. And with the probabilistic forecast, to add a two percent risk that is a bit guesstimated, to say a major drop of demand, that’s technically fairly straightforward.
In contrast, if you’re doing it, if you approach the future with classical deterministic times series forecast, it’s almost impossible to do that. Yes, you can say we have a scenario where there is a disaster happening, but how do you reconcile this scenario, which is widely divergent from your primary forecast, with what you’re doing on a daily basis? It’s in practice, you can’t.
So there are many companies that say, “Oh, we do scenarios, we model risk,” but the reality is, what about your day-to-day decisions? All those day-to-day decisions are 100% driven by the median forecast or the average forecast, which completely ignores all the risk. So in this sense, yes, you did some intellectual exercises to think about risk, but if all the decisions that you take on a daily basis do not embed this risk one way or another, then this is just an intellectual exercise. It has no consequence on what you’re doing on a daily basis.
Conor Doherty: I do want to push you a little bit on this point because I actually am also curious. If you think of other forecasting techniques like, let’s say, forecast value added in which you have people collaboratively adding to a forecast and the idea is that different departments have insights. Take an example, a new competitor is about to emerge and you take that information, marketing has that information and they sort of somehow fold that into a time series. That’s kind of difficult to do because how do you translate that sort of knowledge into a forecast? Similarly, this is where I’m pushing a little bit, how exactly does one factor a two percent chance of war in Western Europe into a probabilistic forecast to arrive at the number of units I have on my shelf? Because they seem kind of similar in a way.
Joannes Vermorel: Let’s start with the time series. You see, time series are nowadays people think that there is this general belief, not everybody, but most mainstream supply chain practitioners only think of the future through the lenses of the time series. Time series are incredibly narrow as a way to express anything that you know about the future. For example, if you’re a B2B company, so your clients are other businesses, a very basic risk is just one of those big clients leaves you to go for another one of your competitors. And when this happens, all the products that they were buying from you, they would cease to buy anything. And if you had, for example, you were keeping in stock a product that was very routinely purchased by this one customer, but this one customer just suddenly leaves you, then this pile of inventory just becomes dead stock overnight. Just because, although the stock was rotating nicely, it came with a hidden risk that this client could leave you.
So here we have, and the idea that these big clients can leave you is not like a super sophisticated idea. Any salesperson would say, “Well, we had this client, there is always this risk that they leave us.” Now the problem is that if you frame your anticipation of the future with time series, you’re stuck. You cannot express that because the information that you have is about the client, not about the products. And if you said there is this risk of this product going to zero, yes, but the thing is this risk is highly correlated. It’s all the stuff that this client is buying that can go to zero at the same time. And it’s a very, very different sort of risk of saying that this product in isolation can go to zero.
The first thing is that time series are just not appropriate to even express risk.
To ensure people didn’t miss that, time series is a reflection of a client’s relationship to a product, but not the products themselves. Time series is just a one-dimensional measurement. You have a measurement that falls every day, every week, every month. That’s called equis based time series. That’s what people have in mind when they think time series. It’s a one-dimensional measurement and it’s literally like temperatures. There were temperatures in the past, there will be temperatures in the future, and so you can extend this time series.
However, this is about previous relationships that pre-existing customers had with the products you bought, but that says nothing about the future. The problem is that the information that you have is at the granularity of the customer, and your forecast is at the granularity of the product. There is a mismatch and there is no translation to go from this information to this other information. That’s a key point.
In mathematics, when you cheat, you end up with weird stuff. For example, when you cheat just a little, let’s say for example the Earth is a sphere, approximately. It’s not exactly a sphere, but it’s close enough. So when you want to have a map, you’re projecting a sphere on a flat surface. If you look at a world map, you end up with distortions. For example, Africa in European Maps appears very small compared to Europe, although Africa is actually bigger than Europe. That’s just an effect of the distortion because you’re using a flat surface to represent a sphere.
But here, the problem is vastly bigger. You’re trying to represent something totally risky. It’s a many-dimensional object that you’re trying to represent as a one-dimensional object, your time series. So the sort of problems that you have and distortions that you have are absolutely gigantic. If you think that just making Africa look smaller than Europe is a problem, those are very modest problems compared to the sort of problems that you have in supply chain when you try to inject those information that you know about the risk into the time series.
We have another problem. When you don’t know the solution, it’s very difficult to think about the problem. People are not really familiar with the class of mathematical models that could represent those risks. They are stuck with time series due to the fact that they can’t even imagine something that would be anything but a time series. But the first step is to acknowledge that this is not a correct representation. It doesn’t matter if it’s not very clear yet what should be used.
There are some technical things. For example, it’s not very clear how a logarithm is computed, but that’s fine. You don’t necessarily need to have a clear picture of the thing to use it successfully. So then we can go to the second part, how does Lokad leverage this sort of information.
The idea is that when you want to think about the future, the high-dimensional version is to think that every single possible future has a probability. So you could think of it as this probability for any given future where you know exactly the sales level of everything, about demand, things that will be sold again.
There is a probability of this happening. It’s vanishingly small, but if you have the proper mathematical tools, you can work with vanishingly small probabilities. And again, due to the fact that you have a very large number of possible futures, it will still add up to probability one. There is one future that will happen, and the sum of all those probabilities equals one.
You can factor a risk like two percent chances of losing these clients. It’s not actually that difficult. If you see the demand through the lenses of the products, then it’s very difficult to inject the clients. But if you see the demand as the results of the behavior of the clients and you forecast the behavior of the clients, then adding this extra risk of the client leaving you becomes something relatively straightforward.
You can build your forecast in different ways. In terms of the agility that one has using or leveraging a time series approach versus the agility one has leveraging the probabilistic approach, what’s the difference there and how does it then translate to managing risk?
The main problem is that time series risk does not exist. They can’t even exist. It’s like a cube in a two-dimensional space. There is no such thing as a cube. You can draw a cube, but fundamentally, it just does not fit. That’s a problem when you have extra dimensions that do not fit, you’re stuck. If all you have is a two-dimensional plane, you can’t stack a cube in that. It’s just not going to fit. And so with time series, you’re kind of stuck.
You could duct tape things. You could say we can’t deal with the risk, but we can cheat by having an incorrect forecast that is intentionally distorted so that the decision that is going to be made based on this forecast reflects this risk. That’s a very convoluted way to get to risk management.
Technically, it is possible to kind of do it, but it’s going to be in ways that are very strange. For example, you can deal with the risk by making your forecast intentionally less accurate by introducing distortion, intentional distortion in your forecast. That’s one way to deal with risks. But that’s a very convoluted way to get to that.
If you go to the probabilistic approach, you have an inherently probabilistic forecasting. Then by design, you have those probabilities. There is another part of the challenge which is how do you do an optimization. It’s called the stochastic optimization process. How do you optimize a decision when you have uncertain conditions? So you need to do an optimization that has a natural affinity to these uncertainties that exist in the initial conditions.
Conor Doherty: If assessing, and I’m curious how exactly, well no, let me restart that question. If you’re in a situation where you have a company and you embrace the probabilistic approach, you have been doing time series, you’re won over by what you just said and then they’re presented with a recommendation which is the ultimate product of the probabilistic forecasting methodology. And in that, there in whatever value has been presented, there’s actually a lot of those factors have been baked in, for example, the possibility of losing a client and you know the management look at that and think that’s insane. How exactly are they supposed to interact with that because again there are so many things being factored into that. How do you bridge the gap?
Joannes Vermorel: So first, what is the output and that’s where there is a radical divergence. The output of a risk-driven supply chain process that is powered by probabilistic forecasts, because literally, it is to my knowledge pretty much the only viable technique that we have to deal with risk. That’s what probabilities are for. The output is the decisions, not the plan. That’s a weird thing is that when you think that the future can be known so you can eliminate all the risk, supply chain risks are mostly about this uncertain future. If you think that you can have an accurate forecast then the output of your supply chain practice is the forecast and the forecast is your plan because once you have the forecast it’s just a matter of orchestration for the decisions.
If you take a risk-driven approach then the output of your process is not the plan, it’s not the forecast, it’s the decisions. But if your risk-driven process is bad because it can be bad, it will lead to bad decisions. And how do you identify so how do you challenge a decision to be bad? Well, it happens very differently. Again, if we go to the classic perspective, people would think in terms of forecasting accuracy because that’s the end game. If you go for risk-driven, you would say there is a decision, this decision has risk and opportunities attached to it expressed in dollars or euros. And so if you see a decision that is bad, essentially you’re saying the assessment in dollars or Euros that has been made about this upcoming decision is wrong.
And so you can pinpoint and if you look at it, we would typically for every decision that we generate, we decompose the economic drivers so that we can say we have half a dozen of drivers that reflect what goes into this decision. And so if you want to challenge that, you will challenge a component and say this, let’s say the carrying cost, the risk of carrying costs you estimate looks completely off. And so yes, that’s the role of the supply chain scientists to reverse engineer the process to identify what’s wrong with this estimate. But it is very technical.
But the reality is that if you have a classic time series forecast that is very wrong, you say this time series forecast is very inaccurate. But once you say that, investigating the root cause of that is going to be a very technical undertaking as well.
Conor Doherty: If we go back to earlier, we were talking about proactive approaches to risk management which was let’s say the digital twins in the automotive industry and then to sort of reactive risk management from the clothing analogy you gave. Probabilistic forecasting sounds almost proactive in the sense that you’re simulating worlds in which you make this decision, here is the anticipated response, you make this decision, here’s the anticipated response.
Joannes Vermorel: So it is proactive in the sense that you just say there will be fluctuation, there will always be fluctuations that are way beyond my control. That’s what this irreducible uncertainty about the future is about and thus based on that I need to be able to engineer a process that is going to react swiftly and adequately to those changing conditions whether they are impacting me positively or negatively. And so yes, it is very proactive in the sense that engineering such a process that lets you take advantage of the opportunities as they arise and mitigate the problems as they arise, it takes a lot of preparation.
But it is not fooled by the idea that you can prepare so much as to lead you to an elimination of the uncertainty in the first place. You see, that’s not the end game. It’s a bit of a dogmatic position in a way but it’s the idea that you can’t get to the bottom of this predictive modeling rabbit hole. You can’t get to a model that is going to be 100% accurate, that’s never going to happen. The amount of residual uncertainty is going to be very large and thus what you’re left with is engineering a process that is very good at keeping up with the change as you observe it.
Because you see, the thing is that very frequently companies get surprised by stuff that happened months in the past. You would say oh we don’t know the future but what about the past, we know the past. But if your average response time to something that you’ve already seen is like six months, then you may end up being surprised by something that is already a couple of months in the past and companies very regularly do get surprised in this way.
Conor Doherty: Again, I do want to push a little bit here because I know if we want to talk about risk management, we should talk about how we assess our risk management practices. And again, to come back to time series versus probabilistic approaches, if you have a time series and it’s wildly wrong, I can point to that and say well that was wrong, it was massively incorrect. And that is, you know, it’s binary, it either was accurate or it was not. You said we would sell 100, we sold 10. You were out by an order of magnitude. With the probabilistic approach, you’re providing probabilities, you’re not saying this is definitely what you’d sell. And does that sort of insulate you from being wrong?
Joannes Vermorel: No, I mean, technically there are metrics for the accuracy of probabilistic forecasts but even more interestingly, you can assess the correctness of the decisions themselves. And that’s, you know, forget the probabilities. They are just a transient computation artifact. There are plenty of other artifacts, numerical artifacts that go into the calculation. They are inconsequential in the sense that if you have the incorrect probabilities but you still get the correct decision, does it really matter that your probabilities are wrong?
Conor Doherty: What do you mean by that? You could have the incorrect probability but you still make the right decision?
Joannes Vermorel: For example, there are people who don’t always realize but computers approximate stuff all the time. Whenever you do a calculation, you just use a certain number of digits of precision. Is the loss of precision important or not? The answer is, it depends. And in the supply chain, it depends on what. Well, it depends on whether the final decision is good or bad.
So in the end, what I’m just saying is that you should judge the quality of this risk-driven approach by what it does at the very end of the process, the decisions. Dealing with high dimensional probabilities, to do with numerical assessment of probabilities in a very high dimensional space comes with all sorts of quirks. Whether the techniques are appropriate or not should really be judged by the end game results, not by the quality of a probabilistic forecast.
Necessarily, the accuracy of the forecast isn’t the main concern, but rather how much money was made or lost.
Conor Doherty: Yes, exactly. And well, that’s very difficult for some people. Forgive me, I don’t want to be condescending, but are you saying the idea of wanting a more accurate forecast is technically wrong in terms of risk management?
Joannes Vermorel: So, first, I say when you say you have, let’s say, a 20% inaccurate forecast, those percentages are a completely made-up unit. They are not kilograms, they are not kilowatts, they are not something that has any kind of tangible reality. This is made up and people say, “Oh, but we’re so used to those accuracy percentages being expressed as a percentage that surely it must be real.” I say not at all. You can have exceedingly damaging forecasts that happen to be very accurate, where the inaccuracy expressed as a percentage is very low.
There’s an anecdote that I’ve been given over and over where you can just forecast zero demand for a store and that will very quickly give you a very accurate forecast. You forecast zero, you open with zero, and the forecast becomes 100% accurate. So, this measurement expressing percentages is not very sensible.
If I tell you that you can have a probabilistic forecast measurement expressed in cross entropy, that’s very abstract and not going to be super insightful. But the case that I’m making is that cross entropy is as abstract and opaque as the percentages. It’s very made up. The only reason why, for example, at Lokad we would pick cross entropy is that it has good properties when it comes to getting to the final decisions.
For example, cross entropy exhibits very steep gradients which facilitate learning high-quality models. That’s a very technical thing, but it works. And it works in what sense? It works by judging the end game results, which is the decision generated at the very end of the process and which ultimately results in reduced Euros or dollars of error. That’s the metric that’s salient for people who operate from this risk-driven perspective.
Again, if you operate from the time series perspective, you’re thinking in terms of just like a car manufacturer with a defect that would kill people. You say, “You know what, we don’t count dollars, we just want to make sure that we are exceedingly safe and that we are safe almost beyond measurements.”
Conor Doherty: So, if you have an entire division dedicated to managing risk, assessing risk, but your supply chain is predicated upon a time series forecasting approach, is it your position that that’s almost paradoxical, like it’s a contradiction in terms?
Joannes Vermorel: No, it just means that the people who are doing risk management are just bureaucrats. Whatever they do has no consequences. Usually, they just have no consequences. You see, the thing is, if you make a risk assessment but if these things cannot marginally adjust every single decision being made in your supply chain, then you’ve made an assessment and you’ve just buried the assessment right after making it.
You see, if you say, “Oh, this supplier has a 2% risk of going bankrupt next year,” okay, does it impact your purchase decisions? If it doesn’t, then you’ve just buried your assessment. You’re ignoring it. You’re just sticking your head in the sand.
And that’s very strange because people would say, “Oh, but we have analyzed risk.” Yes, but you’re not acting on this assessment. And when I say acting, people really think about that’s a mistake. When people think about supply chain, that’s what you would hear in the media. They would say, “Oh, we should not have our factory in China.” Yes, that’s a very macro risk, but there are also much more mundane risks.
So, what do you buy, where do you stock it, do you increase or decrease your price points? Those are decisions that also come with risk and those decisions are made on a daily basis for every single SKU that you either buy, produce, or sell. And whatever assessment you have in terms of risk about your supplier, your competitor, your clients, the question is, if there is not something that numerically connects the dots between this assessment and those very small decisions that you take, then you’re not properly managing the risk.
Conor Doherty: So, correct me if I’m wrong, are you saying that most people’s conception of risk management is on the macro scale, like a massive event that completely disrupts chains, but your position is that the more important, the more pressing risk management is on the day-to-day, smaller decisions?
Joannes Vermorel: Both are very important, but let’s be real about how much you can be truly informed to make the right decision. For those macro decisions, to a large extent, it’s a gamble. It’s a complete gamble and that’s fine. That’s capitalism. This is an economy of profits and losses. People take risks and there is chance involved. And I’m saying, well, you can’t really have a practice that will tell you whether entering a new market, for example, is safe or not. You can make assessments, you can try to rationalize the process a little bit, but fundamentally, it is something that kind of evades statistics and quantitative analysis.
On the contrary, if you look at a supply chain, a mid-size company is going to make tens of thousands of decisions a day, every single day. And that’s what I’m saying, unlike the big macro decisions where you gamble and there is no other alternative except going with your gut assessment, in the case of those tens of thousands of decisions that have to be taken on a daily basis, you can do a quantitative assessment and something that actually makes sense.
Conor Doherty: Well, to get away from the colossal examples like the macro scale ones, let’s bring it down to something like the SKU level. So, we have a number of stores, we have 10 stores, and we have a finite amount of inventory, inventory of white t-shirts, and all 10 stores need white t-shirts. What would be the probabilistic, most risk-informed way of distributing what I have amongst all the stores of need?
Joannes Vermorel: Let’s take the classic perspective. The classic perspective, time series focused, you’re going to assume you know the future. So, you have a safety stock. Basically, you say every store should have this quantity in stock and then to acknowledge the little residual uncertainty on top of that, you’re going to add a small buffer and that’s your safety stock. All stores are treated independently and the idea is that you should have enough stock to cover all your stores.
Now, what is the actual risk? The risk is you can run out of stock at the warehouse level and then the question becomes, I have a resource that is limited at the warehouse, what should I do for my various stores? If you just do it the classical way, the classical way would say I have my safety stock, I just do the allocation for the first store, I still have inventory left, I repeat the process for the second store, and then maybe at the fourth store, I will stop because there is no inventory left. So, what you’ve done is effectively fill up the first four stores and send nothing to the others. This is not very smart. This is not dealing properly with this situation, this small mini crisis that you have by having a product out of stock in the warehouse.
Conor Doherty: What are the risks there though from the financial perspective?
Joannes Vermorel: No, that’s not the case. That’s what your safety stock is about. When you put a safety stock, you’re saying I’m putting units in a store that have a very low probability of being sold during my relevant time window. That’s what the safety stock is there for. It’s a buffer that most likely you won’t need.
If you want to maximize your sale, it’s much better to just spread out the inventory so that every store has a little bit. The goal is for all the stores to run out of stock at the same time. Obviously, you can’t really achieve that, but that’s what you want to get close to.
Let’s consider the alternative situation where you’ve concentrated the stock on the first four stores. You have all the other stores that are out of stock, so you don’t sell at all. And for those, you’re only going to sell something like half of the inventory, so you’re going to have a lot of leftover. You end up in a situation where you have a store that is out of stock while another one has comparatively excess stock and the demand in the stores that don’t have stock goes unserved.
Conor Doherty: So there’s the risk?
Joannes Vermorel: Yes, and that’s where we are talking about the risk of having an out of stock. One way to manage that is to preserve the stock in the warehouse when you see that a product in the warehouse is at risk of going out of stock. This way, the best stores can still have a little bit of merchandise.
As opposed to the time series approach which would treat each of our 10 stores independently, a probabilistic approach will take the network and the contingencies or the interdependencies between all of these stores and the relation to my available inventory into account.
Conor Doherty: I’m curious about how exactly can a company manage all of this because that’s a lot of information compared to let’s say the traditional time series approach. Is the only way to manage all of this through automation or are people still involved checking these decisions?
Joannes Vermorel: The way Lokad does it is by automating the whole process. People are there to supervise the automation, but the reality is that most companies, although they claim that everything is manually validated, have been using fairly automated processes for a long time. Whenever you have a min-max inventory setup with a mean and a max, you have a replenishment automaton that just runs typically unattended. This has already been the case of having extensively automated setups for decades.
Lokad is just one more step in this direction, but it’s not necessarily such a game changer compared to what people had beforehand. It’s more automated, but many companies already operate on highly automated setups.
Conor Doherty: Could a company that doesn’t leverage automation, but let’s say, as I mentioned earlier, has entire divisions of experts in Risk Management, those companies are fairly aware, right?
It also occurs to me that we’ve sort of centered the entire conversation around more tangible risks like skus, stores, floods. These are all very tangible resources or assets and the commensurate risks. Are there intangible risks, things like time, bandwidth, knowledge, all of these things that sort of go into the running of a company? What are the risks there or how do we manage those?
Joannes Vermorel: There are intangible risks. For example, if you are a fashion company and you do sales, you generate an expectation among your customers that those discounts will happen again in the future, and so people modify their behavior. Estimating this process in theory is possible, but in practice, it’s very difficult because building the expectation of your customers is something that is done over many years, so it’s not something where experimentation is easy.
For example, if you’re a luxury brand and you have a conviction that you should never do promotions because it devalues your brand, you’re not going to do a five-year test to see if doing promotion truly devalues your brand. At some point, you need to operate on convictions and judgment as opposed to doing the test.
The cost that is generated by doing a promotion is very real. When you do a promotion, you have a certain amount of money which is what you immediately forfeit by lowering your price, so you give up some margin. That’s an immediate cost, but there is also this extra risk of bad habits being generated on the customer side and you need to quantify those costs.
There are also other classes of risk, such as IT dependencies. You can have software that falls apart, you can have plenty of other things that just impact your supply chain. But these risks are more like those in manufacturing where you want your ERP to be 100% uptime. There is no reason to have downtime, you can engineer your way into something that is incredibly close to 100% uptime.
Conor Doherty: You just mentioned that from a pricing strategy perspective, discounts might inculcate bad consumer habits. What did you mean by that?
Joannes Vermorel: Whenever you do a promotion, the customer sees that you do a promotion. So next time, they will say, “I will not buy full price. I will just wait until you do a promotion again. I’ve seen you do promotions, so I know that promotions do happen with your brand, so I can wait. I will wait until you do a promotion and then I will buy.”
The problem is that nothing will be truly able to model that. Modeling the psychology of your customers is mostly beyond your grasp because it takes a decade to shape the mindset of your customers.
When you set a price, you convey a message to your customers. People pay a degree of attention, but it takes time to sink in. So there is a substantial inertia. You can try to do fancy modeling to estimate exactly what would be the impact, but the reality is that as these sort of things take years, you’re not going to be able to experiment much. You’re not going to be able to validate whatever techniques you’re using. Thus, in reality, you need to make judgment calls.
Conor Doherty: So that’s an element of the risk management protocol that will still be within the remit of people reaching a consensus. Do we want to liquidate that stock? Do we want to hang on to it forever? Or do we want to sell it in a promotion?
Joannes Vermorel: Yes, and when you have this risk-driven approach, it is much more compatible with those sort of guesstimates compared to a traditional time series planning forecast where this sort of things had no place.
Conor Doherty: So the guiding principle for determining which sort of back-of-the-napkin rule of thumb policies companies should take should be, does it contribute to a greater return?
Joannes Vermorel: If something can be quantitatively assessed, then go for it. But when it’s not possible, yet there is a general agreement that it’s important, then you should guesstimate.
I think it’s a very dangerous path to say we don’t have reasonable numbers and thus we pretend it doesn’t exist. It does exist and thus you need to guesstimate. It’s better to have a number that is approximately correct as opposed to exactly right.
Conor Doherty: I feel like we’re kind of wrapping up a little bit, but I would like to ask a bit of a difficult question. You talked about if a quantitative analysis is financially prohibitive. So for bigger companies that can afford more elaborate forecasting and risk management policies, they can perhaps take the probabilistic approach. But for companies that don’t have that sort of disposable cash, what advice would you give in terms of risk management that would be actionable?
Joannes Vermorel: I would challenge, can you really afford to ignore the risk? Inventory costs money. The price plan for having a supply chain scientist to assist you in optimizing your decision is something like 2,500 euros a month. Yes, it’s a substantial amount of money, but if you’re not a tiny company, if you’re a 10 million dollar or Euros plus company, this is not a huge amount of money. It’s actually a fraction of what you pay for just a single person.
If you happen to have five plus people that are fulfilling supply chain functions such as inventory replenishment, production scheduling, inventory allocation, pricing management, and they have a process that just ignores entirely the risk, I would say can you really afford to continue doing that? Ignoring those risks may cost you multiple millions of dollars just because you made a very bad call by completely ignoring the risk.
Due to the fact that it’s unusual, people would expect that it’s only the things that are reserved to companies like Amazon and whatnot. No, it is not. It is, to a large extent, the classical time series forecast is much more complicated and the reason is it’s a mismatched the problem. So yes, on the surface it looks simpler because people are used to time series but when it comes to the actual resolution of the problem it completely mismatched the actual problem and this solution, although it looks easy in practice, is nightmarish to deploy and use in comparison.
The sort of probabilistic forecasts that Lokad is using for small clients, small companies, it is unusual but it does fit the problem nicely and thus in the end, you know, and that’s again in my lectures I give examples that if you want to have a look at probability techniques, most of my code examples are less than 20 lines long. So people would say oh this is impossibly complicated and I say well it’s like 20 lines of code and you can go to give the full detail of the method in a lecture that is like one hour and a half.
Can you really say that something is that your company is so small that you can’t afford spending something like a few tens of hours on the case? Is it really beyond your means? I mean yes, if you’re like a boutique with just one person but if you’re a companies that do 10 million dollar plus of turnover per year, you’re not a boutique. You already have things at stakes and mistakes can cost a lot more and conversely, because it’s not just the mistakes, it’s also opportunities.
If by raising your price at the right moment you increase your margin by ten percent, that can result in a few hundreds of thousands of dollars of pure cash flowing into your company and that really offset the cost of having some people spending time on risk.
Conor Doherty: So if I were to summarize that, it would be essentially there is an element of leap of faith but the water is not that cold once you do…
Joannes Vermorel: I would say it’s not so much a leap of faith. I think there is this very weird idea that from the mainstream supply chain theory, risks do not exist quite literally. You have those super gentle fluctuations of demand and super gentle fluctuations of lead times modeled with normal distributions which is when people say normal distribution, it’s a way to say there is no risk.
The reality is that I’ve never met any entrepreneur that was not fully aware that the business that they have is full of risk all over the place. The sort of insanity is that with the mainstream supply chain software, people pretend that the risks aren’t there but the risks are still there and thus the companies routinely face catastrophes that are very costly due to risk that were not like again I’m not saying like investing entering the Russian Market in 1991 thinking that it will work that suddenly it is going to become the new El Dorado.
I’m saying the companies face catastrophes for things that were entirely preventable, things that were really within the wind of the expected sort of risk such as suppliers having problems, prices going up or down, demand fluctuating but not outside what could be expected from the general evolution of the market. This sort of things and so my point is that there is this insanity where most mainstream practices just completely ignore risk.
When I talk to supply chain practitioners, they would say yes there is plenty of risk but the point is that they can’t bridge a gap and I say it’s not very difficult, it’s just very different from what you’re doing and it’s not only known, it’s actually cheaper because it also leads to a higher degree of automation because also one of the reason why you need so many people when you do supply chain with time series is that due to the fact that you have a massive mismatch between the reality and those time series, you need to have a lot of people to just duct tape the process all the time.
But if you have something where you have a better match, you don’t need nearly as many people to duct tape the thing.
Conor Doherty: On that note, I think I’ll draw things to a close. Joannes, thank you very much as usual, it’s been a pleasure. And thank you very much for watching, we’ll see you next time.