00:00:03 ABC Analysis intro and Pareto principle roots.
00:00:33 Pareto Analysis in social networks, supply chain.
00:02:03 ABC analysis’s role in inventory categorization.
00:03:14 ABC analysis use and misuse in supply chain.
00:06:05 ABC analysis’s relevance in modern inventory systems.
00:08:00 Limitations of the ABC inventory system.
00:09:00 ABC approach’s lack of specificity.
00:11:02 ABC classification stability issues and impact.
00:13:01 Data-driven inventory management proposal.
00:15:38 Forecasting benefits for high-volume inventory items.
00:18:01 Combining economic drivers, probabilistic forecast vs. ABC.
00:19:00 Shift from ABC to informational theory approach.
00:21:39 ABC analysis critique: promoting vanity metrics.
00:25:17 Dangers of ABC analysis reliance.
Joannes Vermorel, founder of Lokad, discusses ABC analysis, an inventory management technique based on Pareto’s 80/20 rule. He explains how it categorizes products into classes according to their sales rates, with ‘A’ representing high-value, fast-selling products, and ‘C’ representing low-value, slow sellers. Vermorel expresses concerns about the method’s oversimplification and potential misuse in modern supply chains, as it fails to capture the nuances within categories. He advocates for a more detailed, granular approach that considers each product’s individual sales history and physical constraints, warning that ABC analysis can lead to misleading metrics and instability due to demand fluctuations and stockouts.
The conversation on Lokad TV is revolving around the concept of ABC analysis, an inventory categorization approach rooted in Pareto’s 80/20 rule. This method is categorizing a catalog based on perceived value and is finding widespread use in ERP software and the broader supply chain industry.
Joannes Vermorel, the founder of Lokad, is providing a detailed explanation of ABC analysis. He is noting that the technique originates from Pareto analysis, which posits that in socially constructed phenomena, the top 20% most significant elements account for 80% of the whole. This principle is observable in various areas, such as social networks, wealth distribution, and product sales.
In the context of supply chains, Vermorel is explaining, the 20% of products that are most important usually constitute 80% of sales. He is attributing this insight to Vilfredo Pareto, an Italian mathematician and civil engineer from the late 19th century. Vermorel is also suggesting that the idea of inventory categorization, even without the mathematical precision, might date back to antiquity.
Vermorel explains that the practical application of ABC analysis for inventory involves categorizing SKUs or products into classes, typically three to five. The ‘A’ class represents fast-moving items, while the last class, ‘D’ or ‘E’, denotes slow-moving items. Classes in between exhibit varying inventory velocities. The result is a coarse-grained categorization that groups products with similar inventory velocities.
Despite the simplicity of ABC analysis, it is widely used in the supply chain industry, although Vermorel suggests it is also widely misused. The categorization, he argues, is beneficial when it accurately reflects the physical constraints of products. For instance, in the aerospace industry, a first category may include expensive, repairable products, while the last category might include unrepairable consumables. In the food industry, separate processing categories might be established for frozen food and dry food. These classifications denote the different methods of handling goods.
Vermorel is outlining how the ABC analysis method originated to manage inventory without having to count every item. Using this system, different categories of products could be stored in different sizes of bins, with a simple rule to reorder stock when the bin reaches half-empty. This approach was practical when there was no software to track and manage inventory. However, Vermorel argues that in today’s digital age, when inventory can be tracked automatically, the ABC method becomes problematic and obsolete.
Vermorel is pointing out that the ABC classification system is a low-resolution approximation of sales history. It does not provide any additional information beyond what could be gleaned from a detailed look at a product’s sales history. As such, decisions based on these classifications could potentially be refined and made more accurate by using actual sales data.
Vermorel also points out that the ABC method can create an illusion of homogeneity within categories. Each category can contain hundreds of products, and there can be significant variations within these categories. For instance, the top 1% of products could have sales velocities ten times higher than the top 10%, and this granularity is lost with the ABC classification.
Vermorel suggests a more nuanced, “divide-and-conquer” approach to inventory management, which places more emphasis on the detailed sales history of individual products. This approach involves assigning a purchasing manager to handle a specific number of items based on their value or importance, thereby ensuring that high-value items receive more attention.
Vermorel is reflecting on the traditional human-driven process of supply chain decision making, which includes decisions about purchasing, production, inventory location, liquidation, and pricing. With the advent of computers, however, this process is undergoing a significant shift. Computers can process thousands of items repeatedly
in a day, depending on what is most sensible, which is a task humans cannot perform manually.
Vermorel explains the concept of service levels in relation to sales volume. He suggests that items with higher sales volume generally have a better forecast because they sell more consistently. An item selling 100 units per day, for example, will likely sell around 100 units the next day, give or take 10%. Conversely, items with lower sales volumes have a more unpredictable demand. They might sell only once a month, and it’s difficult to forecast when that sale might happen.
With a high-volume item, maintaining a high service level is more cost-efficient. This is because the amount of inventory required is determined by lead time and the variability of demand, or “erraticity.” If an item has high erraticity, it requires more stock. However, compared to low-volume items, high-volume items require less stock relative to their sales volume, making them more efficient in transforming currency into extra points of service level.
For high-volume items, a probabilistic forecast would yield a relatively concentrated forecast, reflecting low uncertainty. In contrast, for slow-moving items, the forecast would be more dispersed due to higher uncertainty. Combining this probabilistic forecast with economic drivers, such as the cost of stock, purchasing cost, carrying costs, and the potential margin, results in a more precise service level that doesn’t need crude categorization like the ABC analysis.
Vermorel criticizes ABC analysis as a method that allows avoiding calculations altogether, which was useful in the 19th century when keeping track of thousands of items manually was a daunting task. However, in today’s world, with the advent of powerful computers capable of billions of operations per second, such an approach is outdated.
There are two key reasons for this. First, ABC analysis often leads to vanity metrics, which give an illusion of good performance while the reality might be starkly different. Second, ABC classes are unstable and can change rapidly due to stockouts or demand fluctuations. For instance, an item suffering a massive stockout might drop from Class A to Class B, creating the illusion of an improved service level for Class A, while in reality, it’s just a side effect of poor performance.
Vermorel advocates for a more granular and adaptive approach, which considers as many classes as there are items and possible sales histories. This approach leverages information theory to make better decisions based on observed sales history. In this context, he emphasizes the importance of expressing metrics in terms of dollars of performance or error, rather than in percentages, to better align with the economic realities of supply chain management.
Kieran Chandler: Welcome back to Lokad TV. This week, we’re going to be talking about ABC analysis, an inventory categorization method that has its roots in Perrito’s 80/20 rule. The method works by splitting a catalog based on its perceived worth and has been adopted by many major ERP software’s as well as the supply chain industry as a whole. So Joannes, that’s a quick overview of ABC analysis, but perhaps you could explain it to us in a little bit more detail.
Joannes Vermorel: Yes. As you pointed out, the ABC analysis takes its root in the Perrito analysis, which essentially describes entire categories of social networks. The idea is that the 20% most active or important represent 80% of the mass. This is evident on platforms like LinkedIn, where the top 20% of users probably have 80% of the connections. Similarly, the 20% richest people likely possess 80% of the wealth. In the supply chain context, the top 20% of products account for approximately 80% of the sales. This principle was discovered by an Italian mathematician and civil engineer named Pareto in the late 19th century. As for the categorization of inventory, this idea, albeit without the math, probably dates back to antiquity.
Kieran Chandler: So, how does ABC analysis actually work?
Joannes Vermorel: The key insight in ABC analysis for inventory is to categorize your SKUs or products into classes. You will typically have three to five classes, going from the ‘A’ class, which is dedicated to the fast movers, to the last class ‘D’ or ‘E’ for your slow movers, with classes in between representing varying velocities for your different products. This process results in relatively coarse-grain buckets that gather products with similar inventory velocities, indicating how many units you need to produce or serve on any single day.
Kieran Chandler: Is this method used regularly by companies on a daily basis? It seems very simplistic.
Joannes Vermorel: It is indeed simplistic, but it is widely used, and I would say, widely abused in the supply chain world. This categorization helps when your products reflect the physical constraints you have on your products. For example, in aerospace, you would have a first category of very expensive and repairable products. Then you would have relatively cheap but still repairable products, and then there are consumables which you cannot repair. Each category represents different ways of handling the goods. In the food industry, you would have different processing for frozen food or dry food. These categorizations highlight the distinct physical processes involved. However, ABC analysis itself is not really concerned about the physical aspects of the goods, but only about the sales velocity. The abuse I referred to earlier is the creation of broad categories only defined by relative sales.
Kieran Chandler: Can you explain the typical classification of products? I understand there are classes A, B, and C. What does it mean for a product to be an A item?
Joannes Vermorel: Sure, an A item typically means that you have, say, above 20 units per day being sold or produced. This high turnover rate makes it an A item. This classification system is a numerical cutoff based on the amount of sales over a certain period of time, let’s say the last three months. The exact criteria vary from one company to the next, but the core concept is that you define a velocity as an average over a certain period of time and then set cutoffs for each category: A, B, C, D, E.
Kieran Chandler: It seems logical to categorize products this way, marking some as more important than others. But I get the feeling you don’t completely agree with this approach. Why is it not relevant?
Joannes Vermorel: You’re correct, I have some reservations. The interesting thing about this ABC analysis is that it allows you to manage your inventory without doing any counting whatsoever. Let’s take a 19th-century perspective on inventory management. For Class A items, you would use large bins. If the bin appears half-empty, you place an order. Class B items would use smaller bins, and Class C items might not even have a bin, just a shelf. When one is consumed, you order another. These methods let you manage your inventory without counting anything. You simply pass an order based on the appearance of your inventory, which is great if you don’t have any software to manage your supply chain. However, in today’s world, where we can track and count inventory automatically, this system doesn’t make much sense.
Fundamentally, inventory categorization like ABCD is a low-resolution approximation of how much you need to produce or consume over a period of time. Any decision based on this class could be more accurately made if it were based on the actual sales history of the product. The fallacy is that the ABC class doesn’t add any additional information compared to the raw demand history of the product.
Kieran Chandler: What sort of problems can arise from using this ABC approach?
Joannes Vermorel: There are several issues, primarily stemming from the fact that it’s a very low-resolution system. It’s like having the capacity to do calculations by the gram but rounding everything up by the ton. First, there’s the problem of heterogeneity within a class. If you have 2,000 products and create five classes, each class will still contain about 400 products, and there can be a wide range of variation within those classes. This system gives you an illusion of specificity while overlooking significant differences between products, especially for top sellers, where probably your top 1% of products…
Kieran Chandler: So, when you have the top 10% of your sales, it’s probably operating at sales velocities that are 10 times higher per product. However, if you only have five categories, you’re not going to achieve the necessary granularity. On the other end of the spectrum, for the long tail, it’s very long indeed. Your ‘C’ class, for instance, might contain items that are only sold once a month, as well as items that are sold once every decade. They are packed together, but they are very different. The way you approach something needed once per decade and something needed once per month would be completely different.
Joannes Vermorel: Furthermore, I would say the total problem with a very crude classification is that it tends to be unstable over time. We’ve experimented with several clients and typically noticed that between a third and half of the products change class from one period to the next. When you look at the sales of a quarter, do your ABC analysis, then look at the sales data for the next quarter and redo your ABC analysis, you can end up with a situation where 40% of the products have a different class. That means just because you’ve moved to a new quarter, you’re going to have completely different policies for a given item just because it changed class. It doesn’t really make sense.
For instance, if you had an item that had a gradual but very slow decline over the last two years, why would you decide that just because you crossed a certain numerical threshold from one day to the next, that this item, which you were reordering once a month, you would now reorder only twice a year? That gives you very non-linear effects, with lots of gaps and products that suddenly jump from being reordered once a month to being reordered only once a year. It’s extremely arbitrary and it does not reflect the fine-grained evolution of the demand.
Kieran Chandler: So, how can you approach things in a better way? If you’re not categorizing your products, how do you ensure you’re taking care of those products that are really most important?
Joannes Vermorel: Yes, here it’s typically a divide-and-conquer approach that some supply chain managers have over their own supply chain. They might say, “In order to pay really close attention to the products that matter most, I’ll have a purchasing manager who manages ‘A’ class items, and they will manage 50 items. If it’s a manager that manages ‘B’ class items, this manager will handle 200 items. And if it’s a manager that manages ‘C’ class items, this manager will manage 1000 items.” That way, you have more brain power over each of the items that are most significant.
However, this perspective assumes that all supply chain decisions, like how much to purchase, how much to produce, where to put the inventory, whether to liquidate the inventory, or how to move the price, are driven by humans doing everything manually. But the reality is, as soon as you have computers, everything changes. The computer has no problem processing thousands of items hundreds or even thousands of times a day if that’s what makes sense.
So, if you want to ensure that you have a higher service level for your ‘A’ class items, what you’re really saying is that if you have an item that is selling more, you can have a better forecast. Why? Because it’s a numbers game. It’s much easier to forecast if an item is selling steadily at 100 units a day.
Kieran Chandler: Somewhere around 100, that’s going to be a reasonably good forecast, maybe plus or minus 10 percent. If you have an item that is selling once a month, the most probable forecast for tomorrow is zero units. Yet, it’s very erratic. There might be days where you sell two extra.
Joannes Vermorel: The bottom line is, there’s a silver lining. If you have items that have higher volume, you tend to have lower erraticity. This ensures a higher service level. When we talk about costs, it’s more efficient because the amount of inventory you need to keep is more or less proportional to your lead time on one side, and the erraticity on the other side. If you have an erraticity that’s twice as large, you need to have twice as much stock. You need more stock for items that sell more, but comparatively, to items that have low sales volume, you need a lot less.
Your inventory for high-volume, fast-moving items is more efficient in terms of transforming euros or dollars into extra points of service level. That’s why typically you end up with those buckets and all those recipes. But, you can see the problem completely differently.
First, I would propose adopting a probabilistic forecast that natively reflects the uncertainty. For high-volume items, you will get a forecast that is high, obviously, but also relatively concentrated because your uncertainty is low. This probabilistic forecast provides a distribution that reflects the uncertainties that you will have on future demand.
For slow movers, you will get a probability distribution where the mean of the distribution is much lower because on average you have less demand, but the distribution is going to be very spread because there’s a lot of uncertainty, a lot of erraticity.
Kieran Chandler: So when you say that you want to have a high service level, what do you actually mean?
Joannes Vermorel: What you’re actually saying is that for every stock out that you face, you lose margins. Maybe you have a stock-out penalty, so you lose commerce with your clients when they expect to find something, and you cannot deliver the goods that were basically advertised.
The idea is that if you combine the fact that you have economic drivers, like the cost of stock, the carrying costs, purchasing cost, the cost of stock out, and the margin that is your extra reward, because you’re going to resell your stuff at something that is beyond your cost, then the service level that you get is just a consequence of combining those economic drivers with your probabilistic demand forecast.
The ABC categorization doesn’t even intervene in the calculation because ultimately, the ABC categorization is just a crude way to estimate the future demand or the erraticity of the future demand, which you get out of the box if you have a probabilistic forecasting engine.
Kieran Chandler: So, you’re saying it’s a very crude way of doing it with ABC analysis. Isn’t Lokad’s solution just a much finer grained kind of optimization, kind of an extension of ABC analysis? What’s the key difference?
Joannes Vermorel: The key difference is that we would just consider that we have as many classes as we have items and as many classes as we have possible sales history. If you have a granularity that is so thin that you have one class per situation, the categorization becomes a bit meaningless.
The approach of Lokad is to think in terms of information, more in the sense of informational theory. Where does the information come from to make a better decision? If all the information originates from the observed sales history, then a forecasting model can potentially internally rebuild its own classification.
Kieran Chandler: Classification, if it helps you see, but it sits inside the forecasting model or engine. Typically, because it’s very granular, that’s not how it’s being done. There is no saving. The only benefit of ABC analysis is that it’s a sort of method that lets you avoid doing calculations altogether. This is great if all you have is a pencil, a sheet of paper, and thousands of items. Imagine tracking thousands of items by hand in the 19th century - it’s a nightmare. So, it was very important to have methods where you didn’t need to do any calculations, not even an addition. It’s interesting because these methods were kind of intuitive and represented some sort of estimation approach.
Joannes Vermorel: At Lokad, we take a different approach. Our supply chain approach aims to quantify things to the fullest extent possible. We have incredibly fast computers that can do billions of additions a second, so raw processing power is not a scarce resource. The ABC analysis is simplistic, and as you mentioned, one of the real positives is that you don’t need a pen and paper.
Kieran Chandler: But there must be other benefits of the ABC analysis. Why are companies still using it?
Joannes Vermorel: Well, I’m not entirely sure. We have seen a lot of situations where there are many perceived benefits, but when it comes to real, tangible benefits, it’s much less clear. Companies use ABC analysis as a ‘divide and conquer’ strategy to spread the workload of manually processing the items among many supply chain practitioners. However, the answer is: don’t do that. It’s a bad idea. You should have something that uniformly processes all the items. If your method relies on some sort of categorization, chances are you can rethink your methods to remove that and directly use the average demand over the last few months. This will typically perform better. You do not need something that goes through big increments or big steps.
As for reporting, it’s an approach where management often produces vanity metrics. People are entrenched in the habit of saying, “Look, dear CEO, we are so good when we look at the ‘A’ items because our ‘A’ items have a service level of, let’s say, 95 percent.”
Kieran Chandler: But again, this is a percentage and not dollars of error, so it doesn’t matter if your sales volume, if your items have a very, very high service level. If your clients still think that what they really need is the rest to see it, the thing is that you really want to bring it down to, I would say, dollars of performance, dollars of error. I mean things that are expressed in dollars rather than in percentage.
Joannes Vermorel: And an ABC analysis, it’s typically aware to build vanity metrics where you have percentages and you build extra percentages such as on top of your percentages. And the disconnect with the actual reality can be very, very strong. And if you compound that with the fact that those ABC classes are also very unstable, you know, from one culture to another, you can end up with the illusion that your metrics are good while actually things are changing all the time. For example, if you have an item that is Class A and you suffer a massive stockout, this item is likely to drop into Class B very fast just because the amount of units that have been sold dropped because of your stockout. But then, in terms of statistics, it looks good because you just kicked out from your Class A an item that was badly performing. So suddenly, you know, just because you have items where you have a massive stockout, they will actually exit themselves from the Class A by design because the Class A weighs like the sales volume. So if you have a massive stockout on something that is Class A, it’s not going to last in the class. It’s going to drop very rapidly into Class B.
Kieran Chandler: I see, so you’re saying that if an item experiences a massive stockout, it may appear to be performing well based on the ABC analysis, as it gets moved from Class A to Class B. But this doesn’t necessarily mean that everything is fine because the item performed poorly due to stockout, not because it was genuinely performing well.
Joannes Vermorel: Exactly, all is not well in that case. You just have a side effect where things that perform very badly are exiting your fast mover class, but for all the wrong reasons.
Kieran Chandler: Okay, well, I’m afraid we’re gonna have to wrap things up there. But a big takeaway from today is that ABC analysis is just for the very vain.
Joannes Vermorel: Yes, and don’t do that.
Kieran Chandler: Cool, so that’s everything for this week. Thanks very much for watching, and we’ll see you again next time. Bye for now.