00:00:08 Introduction to Flowcasting and its origins.
00:02:04 Contrast of Flowcasting with traditional supply techniques.
00:04:40 Relevance and challenges of Flowcasting today.
00:06:06 Analysis of Flowcasting’s failure and needed insights.
00:07:06 How stochasticity can cause Flowcasting failures.
00:08:06 Limitations of fractional demand forecasting.
00:10:15 Practical implementation and failures of Flowcasting.
00:11:13 Positive insights from Flowcasting: disaggregated demand data.
00:13:46 Flowcasting’s holistic approach to supply chain optimization.
00:15:14 Difference between Flowcasting and Lokad’s approach.
00:17:31 Complexities of specific supply chain scenarios.
00:19:37 Predictions on future trends in AI implementations.
00:21:43 Value of the Flowcasting book despite flaws.


In the interview, Joannes Vermorel, the founder of Lokad, discusses the strengths and weaknesses of Flowcasting, a method of supply chain planning. He appreciates its focus on demand-centric, disaggregated forecasting, and automation, but critiques its deterministic approach for not considering supply chain randomness. Lokad attempts to manage these complexities with a programming language named Envision, but Vermorel concedes that no system can completely encapsulate the intricacies of the supply chain. Vermorel is also skeptical about the role of AI in supply chain management, drawing a distinction between AI as a trendy term and the real algorithmic techniques. Despite its shortcomings, he acknowledges the value in the insights provided by Flowcasting into retail supply chain management.

Extended Summary

In the interview, host Kieran Chandler and Joannes Vermorel, the founder of Lokad, discuss the technique of Flowcasting within the context of demand-driven supply chain planning. Introduced in a 2006 book of the same name, Flowcasting is seen as the ‘Holy Grail’ of this field, as it promises optimal efficiency by only needing to forecast at the point of sale.

Flowcasting targets retail networks. Vermorel breaks down its core premise as follows: forecast every single product at every single location, each day, and aim for the highest possible accuracy. The comprehensive picture obtained from these forecasts allows for the reconstruction of all necessary supply chain decisions. The flowcasting vision holds valuable insights, emphasizing being as close as possible to demand, a factor that hasn’t been prioritized in older techniques. It also places value in assessing individual product per store, rather than in aggregated clusters, which is a more disaggregated approach compared to the category-wise, top-down forecasting prevalent then.

The implementation of flowcasting, however, proves to be a different story. Vermorel observes that it “never works” as expected. Retail networks that attempt to bring this theoretical concept to fruition find it “brutally nonfunctional.” The root cause is identified as the complete disregard of the concept of “stochasticity” - the inherent randomness in product sales at the store level.

Flowcasting relies on deterministic forecasting, presuming one can predict with absolute certainty the exact quantity of a product to be sold in the future. It fails to accommodate the inherent randomness of real-world retail sales and the fractional demand at the per-product, per-day level. This unrealistic approach renders the forecasts ineffectual. More critically, it leads to the omission of economic drivers, crucial in balancing risks, from supply chain optimization processes, further destabilizing the strategy.

When asked about the practical consequences of implementing Flowcasting, Vermorel admits that, in his experience, retail networks he is acquainted with never really progress past the prototype stage due to the complications outlined. Therefore, while Flowcasting continues to offer some valuable insights, the interview emphasizes the necessity of grappling with its inherent limitations and the realities of supply chain stochasticity to effectively operationalize it in a real-world retail setting.

A critical point of conversation involves ‘flowcasting,’ a well-known supply chain technique. Despite the challenges, Vermorel acknowledges that flowcasting provides valuable insights. He emphasizes the technique’s approach of getting as close as possible to demand at the most disaggregated level. The most disaggregated level, Vermorel asserts, isn’t per product, per store, but rather each unit sold to individual customers. This insight, he posits, gives rise to the concept of time graphs over time series, acknowledging the identities of the customers and their purchasing behaviors.

Furthermore, Vermorel appreciates flowcasting’s emphasis on automation, insisting that Lokad also adopts this strategy. He explains the benefits of daily data refreshes, stating that failing to do so results in unnecessary complexity and missed opportunities, as one would be working with outdated data. This unnecessary complexity, he says, is nonsensical and a waste of resources.

The next insight Vermorel gleans from flowcasting is the idea of synchronicity, and the need for a holistic view of the supply chain. By examining the entire network and all its flows from production to end customer demand, he suggests supply chain issues can be resolved more effectively, rather than simply displacing them. This contrasts with the more common approach of optimizing each stage independently, which may overlook interconnected impacts.

Vermorel’s critique of flowcasting centers on its oversimplification. Although simplicity is generally an admirable goal, he warns that it should not lead to a naive mathematical representation of a complex reality. While flowcasting promotes itself as Excel-simple, Vermorel argues that it oversimplifies the complexities of supply chains.

Lokad’s approach differs from flowcasting in its acceptance of the inherent complexity of supply chains, according to Vermorel. He underscores that supply chains are inherently messy and complicated. Lokad’s model embraces uncertainty, which he views as irreducible. As such, their approach involves a probabilistic algebra, which, while more complicated, leads to more realistic results. Therefore, Lokad’s methodology doesn’t try to eliminate uncertainty but rather acknowledges and works with it.

Further complicating supply chain management, Vermorel details the numerous edge cases that exist, such as differing demands for packaging or specific product formats across stores and individual customers. Such variations, Vermorel argues, require nuanced and flexible solutions that account for these complexities rather than a one-size-fits-all approach.

To address these complexities, Lokad develops a programming language named Envision, designed to maintain simplicity without compromising on the realities of supply chain management. This system provides a better approximation of reality, but Vermorel admits that no system can fully capture the complexities of real-world supply chains.

During the discussion, the concept of Flowcasting, a supply chain methodology, is scrutinized. Vermorel speculates that Flowcasting might have fared better if it had incorporated some form of probabilistic algebra. He points out that despite its shortcomings, Flowcasting was a bold vision for its time.

Shifting focus to emerging technologies, Vermorel expresses skepticism about the use of artificial intelligence (AI) in supply chain management. He contends that AI, while often lauded as a magic solution, will likely disappoint many companies, who will find that impressive technological advances in areas like image recognition don’t translate directly to supply chain problem-solving.

However, Vermorel differentiates between the AI buzzword and fundamental algorithmic techniques, such as deep learning and differentiable programming. These, he says, could be incredibly useful if carefully aligned with the unique challenges of supply chain management.

Returning to Flowcasting, Vermorel asserts that despite its mathematical shortcomings, the underlying book remains a worthwhile read, mainly due to its useful insights into supply chain management in retail. While many of its concepts are yet to be fully utilized, these ideas remain valid and valuable to this day.

Finally, the conversation highlights that the world of supply chain management is a complex one, with no single, straightforward solution. Emphasizing the importance of flexible and adaptable systems that can better approximate the intricate realities of supply chain dynamics, Vermorel concludes by promoting a comprehensive and nuanced approach.

Full Transcript

Kieran Chandler: Today, on Lokad TV, we’re going to learn more about this technique and discuss whether it’s still relevant today. So Joannes, perhaps we should start just by explaining a bit more about what flow casting actually is and how it works?

Joannes Vermorel: Flowcasting was a term coined with the publication of a book in 2006, also titled “Flowcasting.” It’s a technique dedicated to retail networks, offering supply chain optimization for such networks. The book presents many interesting insights and recipes on how to implement it. Simply put, the idea is to pick every single product at every single location, forecast every single day ahead, ideally with very accurate forecasts, and then you get a complete picture of everything that’s going to happen, driven by your client sales. You can then reconstruct all the supply chain decisions you need to make by taking these super disaggregated forecasts that are at the product level per location daily and re-aggregating this data. This allows you to reconstruct all the supply chain decisions by just walking backward from these forecasts, from the store back to the warehouse, back to the suppliers.

Kieran Chandler: Interesting. So, what was so different about this vision? How did it differ from some of the techniques that were around at the time?

Joannes Vermorel: It differed in its series of very acute insights, particularly the importance of being as close as possible to the demand. This wasn’t a new discovery, but to my knowledge, it was one of the first books that emphasized the need to be super close to the best demand signal you have: your sales at the very end of the chain. This differs from many supply chains where, typically, one echelon only looks at what’s happening at the next echelon. For instance, as a supplier, you’d look at what’s happening in the warehouse that purchases from you, but you wouldn’t look at what’s happening in the stores. Flowcasting, in this respect, was very coherent. It advocated for getting close to the demand signal and suggested there was value in looking at the most disaggregated level - product per store - instead of just re-aggregating everything per week per product, just because it’s numerically easier to process.

Kieran Chandler: So, by “disaggregated,” you mean breaking this down to every single SKU level in every single store?

Joannes Vermorel: Exactly. That’s something that Flowcasting suggests and differs from most of the techniques of the time. Typically, people were doing top-down forecasts, where you forecast by categories, split it by subcategories, then by regions, and so on. At the very end, you may have a couple of clusters of typical stores and say, “well if I have store type A, type B, type C, I’m just going to allocate based on the store profile,” and so on. This method avoids dealing with the most disaggregated stage, which is every single product at every single location, where the amount of data just explodes.

Kieran Chandler: So, why are we talking about this today? Aren’t there more modern techniques that have replaced this? Why is it still relevant?

Joannes Vermorel: Well, ten years ago, Flowcasting was all the rage, and I’ve seen at least half a dozen large retail networks get very interested in its insights. However, it never worked in practice. It was like a book full of seemingly excellent ideas, but when people tried to implement it, it failed entirely. It was brutally nonfunctional, which was unexpected given how simple the ideas seemed. The book made it look like you could easily do it with daily forecasts, and it gave some simple recipes to produce those forecasts and to leverage them afterward for supply chain decisions. It turned out that when you tried to put it in practice, it violently didn’t work. The reason Flowcasting failed is that it missed some supercritical stuff. However, most of the market and most actors are still missing those insights. Flowcasting has gone out of fashion after a decade, but the reasons for its failures are still not well known. This means that unless many actors upgrade their understanding, they’ll face the very same problems.

Kieran Chandler: So, on the surface of it, looking at a point of sales, forecasting there, and having it flow back throughout your whole supply chain seems fairly logical. So, where does it all fall apart then?

Joannes Vermorel: It falls apart due to stochasticity or uncertainty. Sales at the store level are highly erratic and random. Yes, you might be selling about one unit a week, but you never know which day of the week. Flowcasting relied on a deterministic forecast where you could say with perfect accuracy that you’re going to sell exactly one bottle of shampoo two days from now for a given reference. However, that’s not realistic due to the inherent randomness in retail sales.

Kieran Chandler: And that’s where everything actually falls apart. Those forecasts have no way to reflect any kind of uncertainty. That’s where forecasting completely falls apart — uncertainty is absent. Thus, because uncertainty is absent, and you assume that your forecasts are accurate, your forecasts fail to accurately reflect the future. They are not probative forecasts. When it comes to supply chain optimization, you act as if your forecasts are correct, and so you ignore what we call economic drivers, balancing all those risks. Essentially, you have a problem that ignores uncertainty, and as a consequence, you also ignore the economic drivers that are central to optimizing under uncertainty. So, if there were so many problems with this approach, what actually happened when retailers tried to implement it? Were there any widespread disasters?

Joannes Vermorel: The retail networks that I was in touch with, they never really got past the prototype stage. They were cautious, they tried it on a very small scale. But actually, the numbers that came out were so nonsensical that supply chain practitioners were saying, no, we can’t push that to the store. It was either way too less, way too few, or way too much. It was literally not making sense. To my knowledge, no large-scale retail network went into even the prototype stage that was more than just a few dozens of products in a few stores, something very small scale, and terminated very quickly because it was dysfunctioning brutally.

Kieran Chandler: So it can’t all be bad then. Flowcasting was a fairly well-known technique, had a good reputation. What are some of the good insights that were given to us?

Joannes Vermorel: There were several interesting insights. They suggested getting as close as possible to the demand and that the most disaggregated level is per store, per product. I agree with the idea that you need to be as disaggregated as possible, but I disagree that per product, per store is the most disaggregated level. The most disaggregated level is directly every single unit sold to every single customer. Nowadays, most retail networks have customer tracking through loyalty programs. So you know not only that you’ve sold this product in this location on this day, but you also know to whom you’ve sold this product. So you should think not just in terms of time series, as described in flowcasting, but as a time graph where you know exactly who you are selling to. That contains even more information. However, the key insight of flowcasting about getting at the most disaggregated level is profoundly correct.

Another aspect they got right was the idea of automating everything with daily refreshes. That’s exactly what Lokad suggests till now. If you do not have complete automation — getting the data from the last day, crunching it, and refreshing all your forecasts and decisions based on your forecast — then you’re leaving a lot of money on the table. Just because if you’re using data from last week, you’re trying to forecast yesterday’s sales. But you don’t need to forecast yesterday’s sales. You already know them. So, having those daily refreshes is a good practice, otherwise, you end up using data from last week to predict what happened yesterday, which is nonsensical.

Another good insight from flowcasting is what they call synchronicity. They express the idea that supply chains need a holistic vision across your echelons. That’s what we did with the Bridgestone case that we discussed in a previous episode. If you want to optimize your supply chain, you need to have a holistic perspective on the whole network and all the flows, from the end customer demand back to the production. If you do not have this kind of holistic perspective, then you’re just displacing problems instead of solving them. So, Flowcasting had this very correct insight that you need to embrace the network as a whole, instead of doing stage optimization, where you optimize the warehouse without considering anything but the warehouse.

Kieran Chandler: So what you’re actually saying is that Flowcasting had quite a few valuable insights that we share here at Lokad. So what are the key differences between the Flowcasting approach and what we do here at Lokad? What are the things that we do much better in your eyes?

Joannes Vermorel: One of the key tenets of Flowcasting was that it can be made extremely simple, Excel simple, with very simple formulas that can be combined in Excel. Simplicity is very good. However, when you model a complex system, such as supply chains, you must ensure that your model is not simplistic. Your model should not betray the realities that you’re trying to represent by having a very naive mathematical representation of a reality that is not so simple.

At Lokad, we’ve tried to really embrace the fact that reality in supply chains is very messy and often complicated. That means that uncertainty is here to stay. If you say, “we just need forecasts that are accurate,” that’s wishful thinking. There is an irreducible amount of uncertainty. You need a system that embraces the fact that uncertainty is here to stay. You end up with a probabilistic algebra, which is a lot more complicated. But it’s what it takes; otherwise, your results are nonsensical.

And if you want to extract as much information as possible from the most disaggregated level, which are the time graph of even clients, who is buying which product, where and when, then suddenly you can’t extract all the information you should from this data with just time series. Yes, it’s a time graph, and it’s more complicated, but that’s the reality.

Kieran Chandler: You have a lot more information in this time graph. That’s what you want to leverage. So maybe another perspective is instead of trying to sell simplicity, you want to have a system that is as simple as possible, but never simpler than what reality actually requires.

Joannes Vermorel: Yes, indeed. I could elaborate on the fact that there are plenty of edge cases, particularly challenging ones, in supply chains. For instance, consider a warehouse that ships packs of bottles to stores, yet in the stores, it’s the individual bottles that get sold because customers sometimes unpack the bottles. And then you have cases where some customers specifically want to buy a pack of bottles. You cannot just count the bottles and assume that’s the solution. Some clients are very picky and insist on having a pack of six bottles. If that’s not available, they prefer buying elsewhere. These things are quite tricky.

I’m afraid that there’s no simple silver bullet approach to this. At Lokad, we realized this fact and decided to create a tool, a programming language called Envision. We designed it to be as simple as possible, but without oversimplifying the realities of supply chains. I must admit, reality in supply chains is really complex. All you can hope for is a better approximation. You will never have the entire real-world modeled in your system.

Kieran Chandler: So, if Flowcasting had some of that probabilistic algebra behind it, it might have worked a bit better in the past, but it was in some ways probably ten years before its time?

Joannes Vermorel: Exactly, I think it was indeed about ten years before its time. Back then, we didn’t have all the technological ingredients necessary to execute what is still a relatively bold vision of the supply chain.

Kieran Chandler: And how about the Flowcasting of 2019? Is there something that looks very good on the surface, but actually doesn’t perform as well once you dive deeper into it? What will be the next thing that is just like Flowcasting that appears promising, but eventually falls apart?

Joannes Vermorel: I believe that most of what is being sold under the banner of AI is going to fall apart. I’m not saying that deep learning or even its descendant, differentiable programming, aren’t excellent tools; they indeed are. However, if all you have is the buzzword and just raw tech coming from domains such as image recognition that are not directly applicable to the supply chain, then many companies will realize that raw tech, even if impressive and can win against a Go champion, doesn’t magically solve their supply chain problems.

I suspect that ten years from now, the AI situation will be even worse than the Flowcasting one. Even with Flowcasting, ten years down the road, there were still some insightful aspects in the book. On the other hand, I anticipate that about 90% of what vendors are currently pushing will become worthless for supply chain purposes in about a decade. They will be completely forgotten. But I’d like to distinguish between the AI buzzword, which I think is just that, a buzzword, and fundamental algorithmic techniques like deep learning and differentiable programming, which can be very useful if carefully aligned with the challenges of your supply chain.

Kieran Chandler: That’s surprising, considering AI is currently a buzzword in the industry. To wrap things up, is Flowcasting still relevant? Is it still worth giving the book a read? What’s the key message to learn from today?

Joannes Vermorel: Yes, I think the book is still worth a read. It’s a pleasant read. The style is nice, and it’s a quick read, taking maybe just a few hours. Many of the insights are still correct. What’s incorrect is the mathematics they illustrate throughout the book. The authors, it seems, were neither mathematicians nor statisticians, and it shows. But if you disregard the math and focus on the insights, you’ll find that they had given serious thought to the supply chain in retail and its implications. That’s where I believe the book really shines. Many of the insights they pointed out ten years ago are still very correct and vastly underutilized. Large retail networks are still not capturing most of the insights described in this Flowcasting book.

Kieran Chandler: Great! Well, we’ve run out of time for today, but thanks for your time. That’s everything for this week. Thanks very much for tuning in and we’ll see you again next time. Bye for now.