00:00:00 Inventory management: Service level, safety stock discussion.
00:00:22 Joannes challenges perceptions of service levels, safety stocks.
00:02:10 Benefits of concrete inventory management decisions.
00:03:07 Exploring service level measurements’ complexities.
00:06:10 Joannes on ABC XYZ analysis’ pros and cons.
00:10:31 Delving into inventory optimization intricacies.
00:11:22 Designing a verifiable system’s complexities.
00:12:15 Critique of ABC XYZ, its psychological roots.
00:13:33 ABC XYZ analysis, human cognition’s influence.
00:16:12 ABC XYZ’s deep dive, computational ranking value.
00:21:04 Discussing inventory categorization and calibration nuances.
00:23:53 Basket perspective introduction, inventory allocation challenges.
00:24:54 Tracing service level history in inventory management.
00:26:55 Pitfalls, misleading implications of service level metrics.
00:28:51 Debunking service level, customer satisfaction myth.
00:32:34 Water analogy for understanding supply chains.
00:34:25 Discussing product sales volume’s dynamic nature.
00:36:00 Quality of service, customer expectations, product availability.
00:38:20 Debunking mathematical trap in product assortment.
00:41:16 Myth of mathematical models in inventory management.
00:42:12 ABC XYZ model’s flaw: Ignoring customer behavior.
00:43:41 Deficiencies of ABC XYZ as prioritization mechanism.
00:44:46 Failed attempts to fix ABC XYZ.
00:47:35 Flawed supply chain assumptions, shift to automation.
00:51:01 Fallacy of average daily sales discussion.
00:52:49 Critiquing product categorization’s volatility.
00:54:07 Disputing mathematical classification’s value.
00:56:11 Deterministic vs probabilistic supply chain approaches.
01:03:43 AI’s usefulness in bridging gaps debated.
01:07:56 Challenging traditional supply chain assumptions.
01:10:33 Tolerance for ambiguity, coexistence of contradictions.
01:16:24 Reality of modern, complex supply chains.
Summary
Conor Doherty and Joannes Vermorel investigate popular stock analysis tool ABC XYZ Analysis, arguing its oversimplification leads to information loss. Vermorel challenges conventional practices of managing service levels and safety stock separately. Vermorel advocates for technology-aided supply chain management, given the complexity of handling large quantities of products. He criticizes the ABC XYZ analysis for lacking dynamism and not considering customers’ perspectives. Vermorel favors a probabilistic approach to supply chain management, which can offer a more nuanced understanding of risks and aid in inventory decision-making.
Extended summary
In this interview, Conor Doherty, the host, and Joannes Vermorel, the founder of Lokad, analyze the popular stock analysis tool, ABC XYZ Analysis. This methodology categorizes products, based on volume and variance, into simple subgroups. Vermorel suggests that this methodology is flawed due to its oversimplification of product characteristics, resulting in a loss of valuable information.
The interview also tackles the complexity of setting appropriate service level and safety stock targets. Vermorel emphasizes the inherent complexity of this task, challenging the conventional idea of breaking down the problem into seemingly simpler parts such as service levels and safety stock.
Vermorel questions the implicit assumption that addressing service levels or safety stock separately simplifies the problem. He suggests that the challenges that arise when determining the appropriate quantity to replenish are the same when figuring out the right service level. The intricacies involved in both processes are similar, making one not simpler than the other.
Explaining his view, Vermorel distinguishes between the tangible, direct decisions about inventory quantities and the abstract concept of service level. He points out that tangible decisions about inventory quantities have clear, measurable consequences on the supply chain, unlike abstract notions of service levels. Consequently, he argues that focusing on tangible, measurable actions rather than abstract concepts could simplify the problem.
Vermorel moves on to critique tools like ABC XYZ analysis, which are used to determine inventory policies. He describes these tools as ‘attention prioritization mechanisms’ aimed at assisting humans in making inventory decisions. While these tools may be helpful in prioritizing which products get attention based on sales volume, Vermorel suggests they do not fundamentally simplify the initial problem.
In fact, Vermorel argues that the focus on developing tools to help humans prioritize attention in decision-making has led us away from the initial problem. This shift, which he likens to the software concept of ‘Yak shaving,’ has resulted in trying to solve a much more complicated issue: how to best present information to humans for decision-making.
He critiques this approach, pointing out that if computers are being used to solve the problem in the first place, there is no need to prioritize the attention of humans. The computer should be allowed to solve the problem in its entirety, without human interference in every step of the process.
Doherty pushes Vermorel on his dismissal of demand variance as a ’tangential concern.’ Vermorel responds by reiterating his core argument: the initial problem was to make the right inventory decision. However, if humans are part of the process, their limited capacity to process information requires prioritization. Tools such as ABC XYZ analysis were created to facilitate this prioritization process, but Vermorel suggests this has led us away from solving the original problem.
Instead, Vermorel proposes that each product be assigned a rank according to its sales volume. The ranking system, he suggests, offers a more informative way to classify products as it retains more data. This system aligns with the computational capabilities of modern computers, allowing for a more precise analysis than the human mind.
Vermorel further criticizes the idea of the human mind being the primary decision-maker in supply chain management. Given the large quantity of products a company deals with daily, he suggests that there is a significant limitation to the human mind’s capacity to manage inventories effectively. He implies that relying on technology to manage these complexities would be more effective.
He discusses the practice of dividing products into categories based on sales volumes, challenging this method as it leads to a loss of information. He compares this method to approximating a circle with a polygon – the more edges you add, the closer it approximates a circle, but it will never be a perfect representation. To Vermorel, the classification of products into a few categories is a crude approximation of the smooth, continuous curve that represents the rank of each product based on sales volume.
Transitioning to the subject of SKUs (Stock Keeping Units), Vermorel argues against treating SKUs in isolation, suggesting that this simplifies the problem but doesn’t resolve it effectively. He critiques the safety stock method, which involves service levels, as it is based on assumptions about future demand and lead time, but these are not normally distributed as the method suggests. He suggests this method can lead to some problematic situations, like negative lead times and sales.
Vermorel posits that the concept of service level is fundamentally flawed. He points out that it may seem intuitive that a higher service level indicates better customer satisfaction. However, the mathematical model underlying safety stock calculations doesn’t provide any insight into customer satisfaction.
He emphasizes the importance of treating supply chain management as a multi-dimensional problem, given the diversity and quantity of products most companies deal with. Vermorel suggests that a different approach should be taken for supply chains with a vast number of SKUs, as the complex, emergent properties of such a system differ fundamentally from those of a simpler, single-product system.
Vermorel then discusses the complexities of supply chain optimization. Just as understanding a single molecule doesn’t provide full knowledge of water in all its forms, understanding one product does not mean understanding the entire supply chain. There is immense diversity and complexity in supply chains, with many elements not conceivable from a single-product perspective.
Vermorel criticizes a common supply chain management approach: the ABC XYZ analysis. He observes that sales volumes aren’t static but rather dynamic, fluctuating widely over time. Even a single product can fall into different categories over its life cycle, making the ABC XYZ model, which perceives sales volume as static, insufficient.
This lack of dynamism is problematic since customers’ expectations are continually shifting, and supply chains must adapt accordingly. If a bakery is expected to have bread available every day, any stockout breaches the ‘social contract’, damaging the client’s perception of service quality. This perception is not determined by the supply chain but rather by the clients themselves.
Interestingly, Vermorel mentions how a single product’s service level doesn’t translate into a satisfactory customer experience when multiple products are involved. For instance, in a supermarket with a 95% service level for each product, a customer who wants 20 products may not find everything, reducing the perceived service level to less than 10%. This discrepancy illustrates the significant difference between mathematical models and customer perceptions.
Vermorel emphasizes that the ABC XYZ analysis, despite its reassuring name (implying safety and control), lacks several important factors. It doesn’t consider the variance over time, overlooks the customer’s perspective, and fails to acknowledge the importance of product combinations in the customer’s shopping basket.
The host, Conor Doherty, adds that if a customer enters a store intending to buy a specific item and cannot find it, they may leave without purchasing anything, leading to a loss of potential sales.
Vermorel critiques the ABC XYZ analysis as an attention-prioritization mechanism, stating that it doesn’t highlight the truly relevant elements in supply chain management. He concedes that the Demand Driven Material Requirements Planning (DDMRP) approach, which prioritizes products based on divergence from target buffers, is more reasonable for attention prioritization.
Vermorel argues ABC XYZ Analysis is not a useful approach to reconcile the complexities of supply chains. He argues that it’s based on a series of flawed premises and that attempts to fix it would only pile up ‘duct tape’ on a method moving in the wrong direction. Instead, he advocates for an approach that appreciates the complexities and dynamism of supply chains and the importance of the customer’s perspective.
Vermorel then delves into the role of technology in supply chain management, pointing out that only recently have machines become capable enough to automate supply chain decisions. This evolution, he remarks, is relatively slow compared to the technological advancements of the modern age. He illustrates this point with a historical analogy: the transition from companies generating their own electricity to buying it from the grid took around 40 years, despite the apparent benefits of the latter.
The conversation shifts to focus on ABC and ABC XYZ approaches to demand patterns, both of which Vermorel finds deficient. He criticizes their static and abstract nature, arguing that they fail to represent real-world phenomena accurately. For instance, he illustrates that product categories can be unstable over time and their classification can jump from one category to another in the ABC analysis, resulting in no substantial value for businesses.
Continuing on this theme, Vermorel criticizes the ABC XYZ matrix as being a mere illusion of pattern, giving businesses a false sense of scientific accuracy when the reality is much more chaotic and nuanced. He argues that these classifications can be arbitrary, leading to an oversimplification of a complex and continuous spectrum of product categories.
The discussion then turns to a probabilistic approach to supply chain management. Vermorel emphasizes the value of probabilistic forecasting as a tool for capturing and processing a significant amount of information, which is useful in assessing uncertainty. This method, he suggests, is especially beneficial because it allows for a more nuanced understanding of risks, further enabling companies to make more informed decisions about inventory quantities.
Vermorel then delves into the role of technology in supply chain management, pointing out that only recently have machines become capable enough to automate supply chain decisions. This evolution, he remarks, is relatively slow compared to the technological advancements of the modern age. He illustrates this point with a historical analogy: the transition from companies generating their own electricity to buying it from the grid took around 40 years, despite the apparent benefits of the latter.
The conversation shifts to focus on ABC and ABC XYZ approaches to demand patterns, both of which Vermorel finds deficient. He criticizes their static and abstract nature, arguing that they fail to represent real-world phenomena accurately. For instance, he illustrates that product categories can be unstable over time and their classification can jump from one category to another in the ABC analysis, resulting in no substantial value for businesses.
Continuing on this theme, Vermorel criticizes the ABC XYZ matrix as being a mere illusion of pattern, giving businesses a false sense of scientific accuracy when the reality is much more chaotic and nuanced. He argues that these classifications can be arbitrary, leading to an oversimplification of a complex and continuous spectrum of product categories.
The discussion then turns to a probabilistic approach to supply chain management. Vermorel emphasizes the value of probabilistic forecasting as a tool for capturing and processing a significant amount of information, which is useful in assessing uncertainty. This method, he suggests, is especially beneficial because it allows for a more nuanced understanding of risks, further enabling companies to make more informed decisions about inventory quantities.
Vermorel highlights two benefits of probabilistic forecasting: it provides more detailed information about the system, and it enables the bridging of financial vision with future anticipation. Unlike point forecasts, probabilistic forecasts lend themselves to numerous methods that can re-express decision quality in terms of euros or dollars.
Vermorel argues that the ABC XYZ forecasting approach represents a dead-end due to its inability to connect metric outcomes with financial results in a sensible way. He criticizes attempts to bridge this gap using artificial intelligence or machine learning, which he compares to attaching an aircraft engine to a slow car. Such solutions, he suggests, are needlessly complicated and overlook fundamental problems that could be more simply and effectively addressed.
The founder of Lokad also emphasizes the importance of quality engineering in supply chain management. He warns against the undue complexity of supply chain systems and encourages focus on solving basic problems. For instance, he cites the hypothetical scenario of a supermarket not stocking a popular brand of diapers, causing customers to walk away, as an issue that won’t be solved by overcomplicated forecasting methods.
Vermorel advises those who are unsure about probabilistic forecasting to challenge their assumptions and question the underlying reasoning of the ABC XYZ method. He contends that while the method does what it’s intended to do (i.e., create a matrix of products aggregated into clusters along two dimensions), the underlying logic and vision of the method are flawed and likely outdated.
Doherty suggests that two seemingly contradictory things can be true simultaneously: an outdated method can work for a time while also not being the best solution. Vermorel expands on this point, implying that businesses often mistake ‘working at all’ for ‘working optimally’. He provides an analogy of carrying water in buckets: while it technically works, better alternatives exist.
Both Doherty and Vermorel agree on the importance of acknowledging the inherent ambiguity in supply chain management and the need for flexibility. The interview ends with Vermorel’s admonition to continually reassess and challenge established supply chain practices.
Full transcript
Conor Doherty: Welcome back to LokadTV. Setting appropriate service level and safety stock targets is tricky, with any number of options on the market and vendors trying to sell you answers. One such tool is ABC XYZ analysis, and here to help me analyze it, is Lokad’s founder Joannes Vermorel. Let’s start right at the beginning - service level, safety stock, all of these inventory policies. Why are they so difficult to set?
Joannes Vermorel: There’s a plurality of options trying to answer these questions. What we perceive as subproblems are not really subproblems. For instance, let’s talk about service levels. There is an implicit assumption that choosing service levels is somehow simpler, like a smaller part of the whole problem. If you can manage that, then you would handle other things as well. The implicit assumption is that we’ve decomposed the problem. The challenge is picking the right quantity for inventory to be produced, stocked, or allocated. When you say ‘service level’ or ‘safety stock’, you are implicitly breaking down the problem. I challenge the idea that this decomposition makes the problem simpler than the original one. When you approach the service level problem, you are facing a challenge that is just as difficult and variable as your starting point. Thus, it’s no surprise if setting a service level isn’t any easier than directly determining the actual quantity to be replenished.
Conor Doherty: So if you could reframe the problem in your own terms, how do you see it?
Joannes Vermorel: In an inventory optimization setting, we are trying to come up with a decision. The decision is tangible. It’s about how many units to allocate, produce, or purchase. This decision will have very tangible consequences on your supply chain. As opposed to, let’s say, deciding to have a 97% service level in this store. That’s an abstraction. There’s no such thing as a 97% service level. It’s potentially a useful artifact, but it’s not something that has a tangible counterpart in your supply chain. When I say it’s an abstraction, I mean that the service level comes with tons of open problems that you don’t have when dealing with a decision. If I decide to allocate 10 units to a store, there is zero ambiguity. I can measure after a while that I decided to allocate 10, and 10 units have been effectively moved. However, it’s not the case with a service level. If more clients show up than I expected, I will not actually get a 97% service level. That’s why I consider it an artifact rather than something tangible reflecting the base reality of your supply chain.
Conor Doherty: And how much of what you just described is actually captured using a tool like ABC XYZ analysis, or its predecessor, ABC?
Joannes Vermorel: Supply chain practitioners want to end up with a decision. If you just look at the numbers and estimate what you need, it’s a very low-tech way to do it. Plenty of stores still work this way. It’s all guesstimation, and it works. However, this method seems crude, so people try to refine it. Then they run into a problem – they have a lot of products, and they realize that the person looking at the product list won’t revisit the case for every single product every day. Thus, we need some sort of attention prioritization mechanism. One way is by sorting the products from highest to lowest sales volume. You can start from the top and work down, deciding to review the top 10 daily, half the list weekly, and the complete list only once a month. That’s one thing you could do, and that’s pretty much the essence of ABC. But the interesting thing about ABC XYZ, is it’s basically a variation on this. It’s an attention prioritization mechanism intended for humans.
Now, at this point, I think we should challenge what problem we are trying to solve. We started with the problem of wanting to pick the right inventory quantity to be allocated, produced, or purchased. That is something very tangible and direct. However, it seems that we went from this problem to another problem, which was kind of picking service level, also picking safety stocks.
Then we entered into yet another problem which is attention prioritization. The pattern that I am seeing start to emerge is something that is, in software, known as Yak shaving. So you were looking to do something very straightforward like, “I want to upgrade Windows 10 to Windows 11.” But then, you end up doing something seemingly unrelated like opening up computers, changing nuts and bolts into the computers. You had a very straightforward goal in mind, but you’ve been sidetracked into doing something only tangentially connected to the original task.
That’s exactly what we are doing here with our inventory optimization problem. We started with a problem that was “Let’s pick the right quantity to be allocated, produced, or purchased.” And now, we are trying to solve a problem that is much more complicated: “How should I actually organize the information to be presented by this human?”
However, this is a very complicated problem. And it is absolutely not clear that solving this problem is the best way to answer our original question. For example, let’s say we have two numbers and we want to add them. Should I really think about designing a system that can present the intermediate steps to a human to verify that the addition is correct? That is orders of magnitude more complicated than just designing a circuit to do the addition.
My criticism here on this ABC XYZ approach is that we started from a problem that was seemingly very complicated. It is actually quite complicated. We tried to decompose this problem, but we’ve been sidetracked. Now, we are trying to figure out another problem that is almost like empirical psychology: How to organize the proper attention prioritization for humans. But if you’re going to use a computer to solve this problem in the first place, why do you need to prioritize the attention of the human? Just have the computer solve the problem for you.
Conor Doherty: If I can push you a little bit on that, because I followed, but as a proxy for the audience, I understand ABC analysis is based generally speaking on either sales volume or sales revenue. We decompose our SKUs into three categories: A, B, C. XYZ is a second dimension, generally demand variance. And if I understood correctly, you were essentially dismissing the quantification of demand variance as a tangential concern. Could you explain why?
Joannes Vermorel: We started with a problem that was: we want to have the right inventory decision expressed as a quantity. We realized that if we involve a human in the loop, the human has a limited capacity to process information. So, we need to prioritize that. If we just do a basic prioritization from higher sales volume to lower sales volume, we end up with ABC.
Once we have that, we need to support this human operator further by helping them to have a sense of what would be the safety stock and service level that would be appropriate for each one of these lines. But this is just decomposing the problem in a way that is suitable for the human mind to process.
The XYZ is to add another dimension which is going to be about the degree of noise or variation among this list. So we take the first, let’s say ten percent, higher seller of our products, and then we want to split this list into chunks that represent the degree of ambient noise for every product. So instead of having just a list of segments, you have a matrix. That’s ABC XYZ for you.
But this is fundamentally something that is very much engineered as a method for the human mind. The question that you should ask yourself is, if I want a machine to handle the end-to-end process, is there any benefit in this segmentation? Does it help me solve the problem?
Not at all. Critics would presumably point out that by creating, generally speaking, a matrix of nine categories, you can identify the variance and the highest contributing SKUs. Then you can set appropriate levels like how much safety stock do I want for that? What is the level for each SKU? There’s a variation between AX and CZ, for example. Let’s assume for a moment that those two dimensions are informative. Well, from a computer perspective, why would you consider discrete chunks? Why have a half dozen subgroups for the volume and another half dozen for variance? You could just use the ranks so you can rank the products from highest sales volume to least sales volume. You can have a number that gives you the exact rank among your portfolio for the volume. Then you could do the same for the variance.
The ranks give you strictly more information. If you look at your classes in the sense ABC or XYZ, the class is just an approximation of the rank. This approximation serves only one purpose - to be more digestible by the human mind. But from a computer perspective, you just keep the rank. The rank gives you strictly more information. The class is a lossy representation; you lose a lot of information. Nothing good comes out of this loss of information.
If we said those two dimensions are relevant, I’m not saying they’re irrelevant. I’m just saying that as far as dimensional decomposition of your problem goes, those dimensions are arbitrary. It’s not very clear that it’s the best way to go at it. If you just look at those two dimensions and preserve the ranks, you will have something that, as indicators, will create a pair of ranks for every single product. This pair of ranks is strictly more informative than your pair of classes.
It’s not just a method that comes with the volume and the variance being of interest; it is very much engineered from the start to have the human mind as the processor of this information. And that’s where I challenge - why would you want that in the first place? We have super powerful computers. Do you think that there’s something that requires the human soul to take those inventory decisions?
If we look at a store that has 10,000 products, all those things are rotating every day. Do you think that there is something for the person that is going to spend, on average, something like four seconds per product? Will there be something like a spark of genius injected into that?
I’m not challenging that the human mind can do incredible things when given time and resources. If you take an Albert Einstein and give him months or years, he can do things that are incredible, way beyond what we can do with machines. But this is not the context in which we operate in the supply chain. People are under immense pressure to just get things done.
And so, if we look at how many seconds of brain power you will be able to allocate per SKU, usually it’s very little. For most industries, it’s going to be a matter of seconds per SKU per day. We’ve discussed the categories, but we haven’t discussed how the categories are calibrated. That’s the result of multiple human minds, as far as I understand.
But if you see that you can have the ranks and now you can decide with percentiles that you’re going to have a split, you can say category A is up to the percentile 10. It’s top 10 or percentile A is top two percent because when you plot all the products from the most sold to the least sold, what you get is almost invariably a Zipf curve, as I’ve touched on in one of my lectures. This curve is continuous, with no plateaus or discrete splits, it’s completely smooth.
It’s akin to approximating a circle in old video games where you had to approximate the circle with a polygon. If you did an octagon, you would get a low-resolution circle. By adding more edges, you get closer to a circle visually. If you have thousands of edges, you get something that looks very much like a circle.
But what I’m seeing here is as if you’re trying to approximate a circle with a square. If you have four classes, you’re approximating your segment with a square. If you have five, you will have a pentagon and so forth. The more classes you add, the better your approximation. But if you remove the approximation altogether, you’re left with the rank of every single product.
So, I’d say, don’t introduce groups, stick with the ranks. If you assume that the volume and variance are useful dimensions, which I challenge, then those ranks give you a more informative version of these two dimensions. Any grouping mechanism you introduce will degrade this information.
Conor Doherty: That transitions very smoothly to the basket perspective, which is something that I’m really interested in, in terms of answering this problem. It treats SKUs in combination rather than isolation. How would that fit into this conversation?
Joannes Vermorel: We started with a simple problem, at least simple in its expression: picking the right inventory quantity to allocate, produce, purchase, or save. We’ve been sidetracked by a widely used method involving service level and safety stocks, but I really challenge the validity of these methods.
The service level perspective comes from historically simplistic assumptions about future demand, where we forecast a normally distributed error about demand, same for the lead time. However, the uncertainty isn’t normally distributed, but that’s another issue.
Once we have our normal distribution, which is a Gaussian, we pick a parameter, the quantile that gives the same effect as the service level. That will give me a target quantity that I should maintain for my inventory. This safety stock is a result of the difference between the mean and the quantile when you look at a one-dimensional distribution.
But due to the fact that it is a normal distribution, it goes to infinity in both directions. The classical safety stock model gives you some weird results, such as negative lead times and negative sales, which are very weird but they are part of the model.
This means that you can pick a service level value that can give you any target stock value between plus infinity and minus infinity, depending on how you choose your service level. This is not theoretical, it’s literally what the math tells you. So, whenever you have a Gaussian, you pick your quantile, and that can go to any final cutoff, between minus infinity and plus infinity.
Conor Doherty: Can you explain the concept of service level in supply chain management?
Joannes Vermorel: When considering service levels, it’s crucial to understand that the range can span from negative infinity to positive infinity. Effectively, your service level is identical to the quantity you decide to replenish. For any quantity you choose to replenish, there is a matching service level understood as a normal distribution. It’s not just an analogy; it’s a mathematical equivalence. For every quantity you’re aware of, if you have a safety stock model, there will be a matching service level in this normal distribution setting.
Now, people might be under the illusion that because the service level is expressed as a percentage, it’s simpler or easier. That’s an illusion. The only slightly good thing about it is that it helps to normalize the scale because all your products have varying volumes and viabilities. Expressing your quantity to be allocated, purchased, or produced as a service level target makes it volume independent and bias independent. However, that’s a weak argument.
The term “service level” can be misleading because people may think that a very high service level is always perceived positively by customers. That’s a misunderstanding. The math of the safety stock model doesn’t say anything about customer satisfaction. People tend to think that if they aim for a high service level, it must be good for clients. But that’s a complete non-sequitur.
Conor Doherty: Can you expand more on why this perception of service level could be an issue?
Joannes Vermorel: The issue arises from the naive notion of equating quality of service with a one-dimensional problem. This might have been true in the 18th century for a bakery selling one product, like bread. This one-dimensional perspective still exists in some commodity markets.
But most modern supply chains are dealing with thousands, if not tens of thousands of products. When we multiply the number of SKUs by the number of locations, we can easily get into tens of thousands, hundreds of thousands, or even millions of SKUs for large companies. This significant number of SKUs challenges the one-dimensional analysis.
A difference in magnitude can become a difference in kind. The emergent properties you get when you have tons of products are very different from what you had when you had just one product.
Conor Doherty: When you mention emerging properties, could you provide some expansion? It seems like an important detail.
Joannes Vermorel: Yes, of course. An example of an emergent property is how a molecule of water behaves differently depending on its state – whether it’s a gas, liquid, or solid. If you wanted to explain all the behaviors you can observe with water, it would take weeks or months. It’s not as simple as picking a molecule and explaining it in 30 minutes, which might be possible with high school students. The same principle applies when you’re dealing with a multitude of SKUs in a supply chain, rather than just one. It requires a more complex analysis.
There’s a danger in thinking that once you understand everything about one molecule of water, you know everything about water itself. That’s not quite right. Similarly, when you say, “I have a model that explains one product, and now I can explain my supply chain which is made of many products,” I’d urge caution. There are plenty of things not conceivable in your one product setup. This is just a simplified example that doesn’t reflect the true complexities of your supply chain.
Even if we just consider one product, there are variations over time. For example, if you consider just one product in isolation, its ranking would fluctuate widely over time. Most products have a life cycle where they start slow, ramp up, plateau, and then decline at some point. So, this one-dimensional model, that looks at sales volume as if it were static, is incorrect. It’s dynamic, and that’s another dimension that’s often overlooked.
Part of the quality of service is this dynamic, time-dependent behavior. If we take the example of a bakery, customers expect to find bread every day. Any stock out is a breach of this social contract.
On the contrary, if you’re an unreliable bakery that only has bread one day out of two, but your bread is much cheaper than the competition, customers might still be happy with you. They have a built-in expectation of your service.
The quality of service is not something that’s in your supply chain - it’s fundamentally in the minds of your customers. Not everyone will agree on that, so it’s inconsistent. If we start to aggregate these expectations, it can be misleading.
When we add multiple products into the mix, another dimension comes into play. If customers want multiple products, we have to look at whether they can find a combination that makes sense to them. A common mistake is to assume that if all my products have a 100% service level, then all combinations of products will also have a 100% service level. This is true only if you never run out of stock, which is almost impossible.
When you start examining the probability of availability or unavailability of combinations of products, you end up with a perspective that’s very different from what a simple safety stock/service level model can give you.
Just to illustrate this, let’s take the example of a supermarket that has a 95% service level for all its products, which is quite good. In Europe, there’s an average of 7% stock out on the shelf, so a 95% service level is pretty good. If you have a client who wants 20 products, which isn’t even a large basket typically, the probability that at least one of those products is missing is probably high. I would need to do the math, but assuming independent availability, you likely have less than a 10% chance of finding everything.
So we start with what appears to be very good from a safety stock and demand perspective, giving the impression of a 95% plus service level. But from the client’s perspective, probably less than 10% of the clients who walk into the store will find exactly what they were looking for. These two things can be true at the same time. You can have a 95% plus service level, and yet have less than 10% of your clients who walk away from the store happy.
What about the products that your clients expect that aren’t part of your assortment? Service level is blind in this sense. If there’s a product that is very much in demand, but you just don’t have it, it’s not going to count as a stock-out or a zero percent service level—it’s just not counted at all.
For example, if I go to the extreme and imagine a store filled with products that nobody wants, this store has, by definition, 100% service level. Nobody wants these products, but they are on display, so you have a perfect service level. The more products you have that nobody wants, the better your service level. That’s a completely mechanical problem, a problem with these mathematical models.
We have to be very cautious, especially when these models have positive sounding names like ‘Safety Stock.’ There is a non-sequitur transition where people assume that because it’s a mathematical model that has a good name, it must be good for the clients, but this is an unwarranted leap.
Conor Doherty: To summarize what you’ve said, it’s crucial to understand our critique of ABC XYZ from a basket perspective. Customers don’t tend to buy in isolation. Not having access to a certain SKU can cause them to leave the store without purchasing anything, even the other high-grade items. This means the store loses all potential sales, not just the individual SKU.
Joannes Vermorel: Yes, and if we go back to the original intent, ABC XYZ is supposed to be an attention priority decision mechanism for humans. But is it a good mechanism to prioritize attention? I would say absolutely not. As a prioritization mechanism, it’s poor—it doesn’t highlight anything truly relevant.
And while I’m not a big fan of DDMRP, I concede that as an attention prioritization mechanism, the way DDMRP defines buffers and prioritizes products against the divergence to target buffers makes more sense than ABC XYZ. At least it is halfway decent in this respect. ABC XYZ is not.
Conor Doherty: Is there any way to reconcile ABC XYZ as an attention prioritization tool with these concerns we’ve just described, particularly the basket perspective?
Joannes Vermorel: No, there isn’t. You start with a series of bad premises. First, you say you want to have a human in the loop, which I challenge. Then you make a second mistake with a mono-product, mono-SKU model with a normal distribution assumption built-in. That’s very bad. It leads to catastrophic results. Then, if you make another erroneous assumption of having a discretization of your space, it doesn’t add any information, it actually loses information. We have been sidetracked with tensions that just get worse and worse.
Now we realize that we have plenty of defects that have accumulated. We’re trying to repair these with what could be compared to duct tape, by re-adding variables that give us the ABC XYZ. We could try to find other ways to fix the method, but in reality, we’re going into the wrong direction. Every additional step you take is just adding more duct tape. It’s not good engineering.
The process you’re crafting is simply not very good. Adding more patches won’t make it better. The only solution is to go back and revisit the assumptions that were made. Are they really valid? If not, you should reconsider entirely the approach that you’re taking.
If we return to our starting point, we began with a tangible problem - making decisions for the inventory. But throughout our journey in approaching the problem, we made a lot of assumptions, and now we’re facing the consequences of those mistakes. Once you’ve made a lot of mistakes, you can’t just make a second proof to fix your problem.
This is similar to when you ask a mathematician if a second proof can fix a wrong one. The answer is no. You can’t fix your problem with a second proof. The only way is to discard your incorrect proof, do the work again, and then you can have a correct path. It’s the same with software. If you have incorrect assumptions, you can’t fix them afterwards. You have to go back to the point where you made a mistake, fix it, and then continue on your path.
Many companies have built entire practices under incorrect assumptions. Due to the fact that supply chains are very opaque and complex, people can operate for decades without realizing any better.
It has only been 20 years since we had computing machines that were capable enough to automate the supply chain decisions cheaply. Modern computers, capable enough to deal with the complexity of a modern supply chain, haven’t been around forever. They’ve been around for a relatively long time, but not for centuries. For many large companies that operate supply chains, this automation only became a possibility 20 years ago.
To give you a point of comparison, it took about 40 years in the US and in Europe to go from companies that were producing their own electricity to purchasing electricity from the grid. Adopting a technology can be a slow process. At the end of the 19th century and the beginning of the 20th, both in Europe and in the US, it took about 40 years to transition from generating electricity in-house to buying it from the grid.
So in terms of timescale, the development of machines capable of performing all these calculations without human involvement at every step of the process is still a fairly recent one.
Conor Doherty: Let’s go back a bit. You talked about the static approach of ABC and, by extension, ABC XYZ. Can you expand a bit on both approaches, or any alternative approaches, to demand patterns?
Joannes Vermorel: Well, we are classifying our products according to two dimensions - the mean, or sales volume, and the variance. But these are again abstractions. They’re not real. There’s no such thing as instant sales volume. That doesn’t exist. That’s the difference between tangible decisions, like moving 10 units, and saying, ‘These products, on average, sell 0.5 units a day.’ There is no such thing. The only thing you can say is that over the last two weeks, you have sold about seven units, which approximates to 0.5 units a day.
Conor Doherty: How do you assess this volume in terms of supply chain management?
Joannes Vermorel: This volume and variance are statistical indicators. The question is how stable these are over time. We have done numerous tests at Lokad, and we have seen that for most businesses, even when we just look at ABC analysis, a significant portion of products will change category from one quarter to another. If you go for anything more precise, like per month, the number of products that would change category would increase significantly.
Conor Doherty: So there are issues with this classification method?
Joannes Vermorel: Yes, the problem with classification, especially when you delve into ABC or XYZ analysis, is that you multiply the number of product category shifts. If you double the number of categories, you will see between 80 and 90 percent of the products jump categories from quarter to quarter. This does not provide valuable insights into your business; it’s merely noise.
These indicators were kind of garbage because they create an illusion of pattern. It might seem scientific, but it’s essentially selling an illusion. Organizing products on a matrix may look mathematical, but it’s just arbitrary ranks determined by a committee.
For instance, when you classify people as rich, average, middle class, and poor, you deal with a spectrum that is continuous. Your cutoffs are completely arbitrary. This same problem exists when you classify your products.
Conor Doherty: So, what’s your perspective on a probabilistic approach?
Joannes Vermorel: The probabilistic approach is hard to compare because it’s a complete paradigm shift. The first major difference is whether we need humans in the loop or not. The quantitative supply chain says no. We want to have the very best that modern computing hardware and software can offer for the supply chain. Whether it involves humans or not is relatively accidental.
So, whether supply chain involves humans or not is somewhat incidental. Probabilistic forecasts are very interesting in this regard because they provide an enormous amount of information. We’ve moved from classes, which lose a lot of information, to ranks, which give a spot measurement. But probabilistic forecasts offer a different kind of precision. Instead of a single point indicator, we embrace uncertainty, representing the ambient uncertainty we have about the system. Why does it matter? Computers don’t have the human mind’s bottlenecks and can process huge amounts of information. This method helps collect a lot more information about your system, your supply chain, your products, and more, compared to point indicators.
Yes, that’s one way to look at it from an informational perspective, what you’ve collected in terms of pure information about your situation. Another angle to look at probabilistic forecasting is from a risk management perspective. We ultimately need to bridge our decision into some sort of risk analysis. We are doing all this inventory optimization to decide the inventory quantities we want to allocate, produce, and purchase. The rationale behind these decisions should be in terms of Euros or dollars of error and reward.
Remember, the mission of a company is to be profitable. Yes, there are plenty of other things a company should strive for, but without profit, the company will cease to exist. For companies that operate supply chains, margins are thin, and survival is not a given. Many large companies go bankrupt every year. Therefore, we need to assess decisions in terms of Euros and dollars.
So, probabilistic forecasts provide more information about the system, but they also pave the way for mechanisms that allow you to bridge your financial vision with your anticipation of the future. It enables a richer set of information and provides methods suitable to express the quality of your decisions into Euros and dollars.
On the other hand, methods like ABC XYZ are somewhat of a dead end. They don’t provide an effective way to bridge the gap between these metrics and the desired financial outcome. You can always engineer complex workarounds, but these methods would be better replaced by something that entirely bypasses the ABC XYZ Matrix.
Conor Doherty: Some people argue that you could leverage AI or machine learning to bridge the gap you’ve just described. They suggest that AI could effectively apply a ‘big piece of duct tape’ to ABC XYZ metrics to accomplish what you’re saying.
Joannes Vermorel: You’re implying that we have a method that generates a matrix unsuitable for the purpose, resulting in poor quality input. We then try to connect this with our true goal. However, the input signal is so flawed that we’d need an incredibly sophisticated workaround to bridge this gap. That’s not efficient or effective. Often, people refer to this as a ‘duct tape on steroids’ approach where the aim is to connect something that’s suboptimal to an output and bridge this gap using advanced analytics. This is similar to saying, “My car is too slow, let’s engineer an aircraft engine on top of my car because my car is too slow.” While this might make your car faster, it’s not the right solution. It’s overcomplicated engineering.
If your car isn’t fast enough, maybe you should reconsider whether the engine it has is powerful enough or maybe there’s just too much weight in the car due to things that you’ve put in. The solution shouldn’t always be additive. For instance, bolting an aircraft engine on top of a car to make it faster is not sensible engineering.
Humans have an incredibly hard time connecting the value of these metrics to the associated costs. This often leads to the invocation of analytical superpowers such as AI or machine learning. These are often seen as magic, as if invoking a demigod of data analysis to perform something nearly magical for us.
While there are instances where these advanced methods may work, I would argue that it’s unnecessary complexity. It’s like creating a contraption that’s way too complicated for its own good. Quality engineering is about creating things that are simple and maintainable, not as complicated as you can make them.
If you introduce undue complexity, you might spend more time debugging a super advanced machine learning algorithm that you barely comprehend instead of focusing on basic problems. For example, your supermarket might not carry the brand of diapers that parents want. New parents might walk away from your store because they don’t see the brand they expect, and your service level analysis or AI system won’t tell you that.
Conor Doherty: To wrap things up, what would you say to people who are still advocates for ABC XYZ but are open to being nudged towards the next step?
Joannes Vermorel: I would advise them to revisit their assumptions and challenge their visions that are going into their requirements. Don’t be fooled by the tradition argument. Just because something has been done for decades doesn’t mean it’s still relevant. Two centuries ago, the number one job in Paris was carrying water in buckets. That’s obviously not the case today.
When something has been done forever, it probably had some value under certain conditions. It shouldn’t be discarded without careful consideration. But the assumptions underlying the method need to be revisited. When I speak to people who promote ABC XYZ, I encourage them to challenge the assumptions underlying the method. I’m not saying the method is wrong, but rather that the reasoning and vision that underline the method might be flawed or obsolete. That is what you should be focusing on.
Conor Doherty: Well, if I can add one little thought to the end of that, I would say personally in terms of ambiguity tolerance, two seemingly contradictory things can be true simultaneously. For example, perhaps you have used ABC or ABC XYZ for decades and it has worked for you. That can be true, but that says nothing about the statement that there are better methods. It doesn’t actually speak to the correctness of the method. So, two things can be true simultaneously and for some people that’s kind of difficult to comprehend.
Joannes Vermorel: I understand that. It’s a confusion of factors and that’s all over the place. Because the reality is, when you say ABC or ABC XYZ has worked for you, I challenge that. ABC XYZ doesn’t give you the final reorder quantities. The problem is that there are other steps after that to get to that and there might be tons of human judgments involved. We started with the idea of having just the store manager that looks at a single spreadsheet, my sales volume, and what do I pick for my products. Then we plug in the middle of this matrix. But if your process is to create a fancy matrix, pretend that you’re a scientist, look smart in front of your colleagues, then discard the matrix and go back to your old ways, you might very well end up saying that it worked very well for you.
It might give your colleague a rationale, it might give you a sort of illusion, a delusion of your own about whether this part of your work was actually contributing to anything. In the end, we were doing something entirely different to come up with the only decision that matters, which is the final inventory decision. Due to the fact that supply chains are very complex and opaque, you can do a lot of things in between that serve no purpose and that seemingly serve a great purpose.
If you look around the world, there are plenty of primitive tribes that will have rituals to invoke the rain. I don’t think there would be many people nowadays that would say that dancing for the rain will impact the weather and improve the yield of your crops. But people would say, “We have danced for the weather for thousands of years, and then there was rain, and then we had a good harvest.”
Yes, you do, but maybe there were steps in what you were doing that were just completely useless. In the end, that’s what you really need to assess. Does this step really contribute as much as you think it does to the quality of the final output, which is a tangible decision, not the sort of artifacts you come up with along the way? Are there alternative methods that would be better? Because ultimately, if you have something that works for you in the sense that it works at all, we are back to carrying water in buckets. It certainly works, but there are alternatives that are enormously better.
Conor Doherty: Well, that’s exactly what I was getting at. The two things can be true simultaneously. You can carry water in a bucket, but at the same time, you could also transfer it in a boat or anything significantly larger is what I’m saying. But again, two things can be true simultaneously, and acknowledging that there is often ambiguity between concepts or the fuzziness that you often talk about can be difficult for people.
Joannes Vermorel: Yes, and this is exactly what you need to change. When people say, “It worked for me,” on these practices that I see in supply chain, you really need to challenge what they mean by “it worked for me”. What does it mean? That’s not a false claim per se, but if all you have to say is “it was kind of right,” it’s not enough.
In a modern, distributed supply chain where your human perception is very limited, you could say the validity of this claim “it worked for me” is absolutely not the same if you’re dealing with a small system on one hand, or a super complex supply chain that you can’t observe in its entirety. Again, if there is a store manager who manages one shelf and says, “You know what, it looks good for me. I look at this shelf and I say this is exactly what my clients want,” I would trust your judgment. That’s because it’s something that is in front of you, you have a sense of the system. You can put yourself in the shoes of your clients. You use your empathy, look at that. You have all the relevant information right in front of you. You can pass a value judgment and this judgment is most likely relatively reasonable, assuming that person is in good faith and whatnot. Now, is this the sort of situation that you face in supply chains?
I would say usually not at all. The typical supply chain situation is, you’re a clerk in an office a thousand kilometers from the place where the goods are going to be shipped and consumed. You’re not looking at the shelf, you’re looking at an Excel spreadsheet. You have dozens of products that you have only seen the product codes for. Most of the time, you have never seen the products for real. And even if you have seen some, you certainly have not seen all of them. You’re serving customers that you’ve never seen, and the data is presented from a system that is super complex and that you barely understand, like your ERP and whatnot. Your rationality is that you’re trying to use your own human rationality to cope with something that is only a tiny, tiny part of the picture.
I very much challenge how much you can say it worked. I could use my own judgment to tell you it worked. You know, if it’s something that is very localized, where you see the whole thing, I would say, “Yeah, maybe you can’t explain to me why it works, but I trust your judgment.” If you’re looking at something that is not even one percent of the total, and you tell me it worked, I say no. You don’t see it, it’s just that it does what you’re used to seeing in this one percent. That’s when you say it worked. You’re just saying what you have in front of your eyes is not deviating compared to what you’ve grown accustomed to seeing for this one percent of the puzzle that you’re looking at.
Conor Doherty: Joannes, I think we’ve covered a tremendous amount of ground today and I don’t have any other questions. Thank you very much for your time and thank you very much for watching. We’ll see you next time.
