00:00:08 Introduction and Stephen Disney’s background.
00:01:26 Overview of the Bullwhip Effect in supply chains.
00:02:26 Four key sources of the Bullwhip Effect.
00:05:17 Stock-outs and their influence on the Bullwhip Effect.
00:06:02 Relevance of the Bullwhip Effect today and ways to mitigate it.
00:08:00 Forecasting and probabilistic models in supply chain management.
00:10:23 Cultivating more options and leveraging substitutions for better service.
00:12:13 Transitioning from point forecasts to probabilistic forecasts to mitigate supply chain issues.
00:14:35 Applying control theory in supply chain replenishment algorithms.
00:15:38 Joannes shares his thoughts on the analogy of feedback controllers in supply chain management.
00:16:00 Stochastic gradient and local optimization in high-dimensional games.
00:17:11 Different optimization approaches for various industries and verticals.
00:19:07 Application of optimization techniques in real-world supply chains.
00:20:54 Importance of accurate forecasting and production engineering.
00:23:20 Relevance of the bullwhip effect today and its relation to the COVID pandemic.
00:25:01 Discussion about the feasibility of quantifying variation in supply chains.
00:26:05 Discussion about financial control in supply chains.
00:26:22 Stephen’s research on dual sourcing supply chains.
00:27:24 Stephen’s views on the benefits of dual sourcing supply chains.
00:28:01 Comparison of supply chains to control systems and natural frequency.
Kieran Chandler interviews Joannes Vermorel, Lokad founder, and Stephen Disney, Operations Management Professor, about the bullwhip effect in supply chains. Disney identifies four key sources of the effect and suggests companies can mitigate it using control engineering ideas, sharing information, and other strategies. Vermorel highlights the importance of probabilistic forecasts and mastering optionality in supply chain management. Disney introduces control theory, and they discuss practical advice for implementing these techniques. Both experts believe the bullwhip effect is not inevitable. Disney’s research focuses on dual sourcing supply chains, offering increased robustness against disruptions. They acknowledge that supply chain recovery from the COVID-19 pandemic will vary based on lead times.
In this interview, Kieran Chandler hosts a discussion with Joannes Vermorel, founder of Lokad, and Stephen Disney, Professor of Operations Management at University of Exeter, about the bullwhip effect in supply chains. The bullwhip effect is a phenomenon where fluctuations observed in a system, such as a supply chain, exceed the magnitude of the fluctuations at the input, typically the demand.
Stephen Disney has spent 25 years studying the bullwhip effect, using engineering techniques, computer simulation, and mathematics to help companies understand and mitigate the issue. He outlines four key sources of the bullwhip effect, as identified in a 1997 paper: 1) demand signal processing (interpreting, forecasting, and generating replenishment orders), 2) batching (minimum or economic order quantities), 3) rationing and gaming (over-ordering due to shortages and canceling orders later), and 4) price variations (manipulating demand for products).
Joannes Vermorel agrees that the bullwhip effect is ubiquitous but notes that the root causes and manifestations may differ between industries. For example, in the retail fresh food industry, stockouts can drive large demand fluctuations by synchronizing consumer consumption patterns.
Stephen believes the conclusions of the 1997 paper are still relevant today, particularly for capital-intensive companies at the bottom of supply chains, as production or distribution can be two to five times more variable than demand, and inventory can be five to ten times more variable. He disagrees with the paper’s claim that the bullwhip effect is inevitable, arguing that companies can mitigate it by selecting appropriate forecasting methods, tuning them, using control engineering ideas, sharing information (e.g., through EPOS data or vendor-managed inventory), and other strategies.
Joannes Vermorel also challenges the idea that the bullwhip effect is inevitable, suggesting that companies can find ways to dampen the impact.
Vermorel highlighted the shift from point forecasts to probabilistic forecasts as a significant improvement in managing supply chain uncertainties. He argued that this shift helps mitigate numerical stability problems and allows for a more accurate representation of demand and lead times.
Vermorel also emphasized the importance of mastering optionality in supply chain management. By considering various options for substituting components or choosing different transportation methods, companies can better adapt to changing situations and minimize risks. He noted that the ability to leverage these options has greatly expanded in recent years, making it increasingly possible to optimize supply chains.
Stephen Disney introduced the concept of control theory in supply chain management, drawing a parallel to the experience of regulating water temperature in a shower. He explained that small, gradual adjustments to supply chain decisions are necessary to avoid oscillations in supply and demand. This concept is applicable to inventory management and replenishment algorithms in ERP systems, where companies can slowly correct inventory levels and work in progress (WIP) to create a smoother and more stable supply chain.
Joannes Vermorel agreed with Disney’s analogy, noting that the breakthrough of deep learning was the rediscovery of stochastic gradient descent, which involves making small adjustments to improve a system. This idea aligns with Disney’s shower analogy, where small, gradual changes can help optimize the supply chain in the face of uncertainties.
They discussed about supply chain optimization and the application of these techniques in real-world scenarios.
Vermorel discusses the effectiveness of gradient-based optimization, stressing the importance of considering the asymmetries of economic drivers in different industries. He uses the example of luxury watch production, where certain constraints may not apply due to the high gross margins and recyclability of materials used. He also emphasizes the need to understand that what may be considered wasteful in one industry might be reasonable in another.
Disney, on the other hand, talks about practical advice for companies in implementing these techniques. He suggests starting with a value stream mapping to understand the process of producing a product and the strategic needs of the supply chain. Companies should determine if they’re focusing on reducing inventory or if capacity costs are significant, as these factors will impact the approach to optimization. They should also examine the time series of demand, forecasts, production targets and completions, and inventory levels to identify the sources and consequences of variability in the system.
Disney recommends considering whether manual forecast adjustments add value compared to algorithmic forecasts and assessing if the chosen forecasting algorithms are properly tuned for the business needs. He also highlights the importance of ensuring the algorithms using the forecasts are set up correctly in IT systems, ERP systems, or spreadsheets for planning production and sourcing from suppliers. Finally, he stresses the importance of production engineering, such as the reliability of machines and product quality.
Regarding the relevance of the bullwhip effect today, Vermorel believes that the COVID-19 pandemic was not a direct manifestation of the bullwhip effect but instead an example of fat-tail events that remind supply chain managers of the importance of considering non-normal statistical distributions. He speculates that the pandemic’s aftermath may give rise to bullwhip-like problems, as seen in the electronics industry in Asia.
Vermorel explains that by examining all possible futures and decisions, companies can achieve more granular responses to supply chain issues. This approach was not technically feasible two decades ago, but it allows businesses to control financial outcomes more effectively in the present. Supply chains often have non-linear costs, meaning that producing twice as much may cost five times more due to overtime, aggressive maintenance, and other factors. While these problems cannot be entirely eliminated, they can be managed more efficiently from a financial perspective.
Disney’s research focuses on dual sourcing supply chains, where companies source most of their products from low-cost countries with long lead times, while supplementing with a smaller local factory. The local factory can respond quickly to demand variability, keeping inventories under tight control, while the majority of demand is satisfied by low-cost products from distant factories. This approach has several benefits, including reduced global distribution, potential for bringing manufacturing back to western countries, and increased supply chain robustness against disruptions.
According to Disney, supply chains have a natural frequency, much like a bridge vibrating in the wind. The supply chain is currently oscillating at its natural frequency due to the impact of COVID-19, which will cause demand to rise and fall periodically. Global supply chains with long lead times will take longer to recover, while shorter lead time supply chains will bounce back more quickly.
Kieran Chandler: Today on Lokad TV, we’re delighted to welcome Stephen Disney, a professor of operations at the University of Exeter, who’s going to explain to us why this effect can occur and what impact it can have on supply chain practitioners. So Stephen, thanks very much for joining us live from Exeter today. As always, we like to learn a little bit more about our guests. So perhaps you could just start off by telling us a little bit about yourself.
Stephen Disney: Yeah, my name’s Stephen Disney. I’m a professor of operations management at Exeter University here in the UK. I’m actually interested in supply chain dynamics for my research area. So, I’ve spent about the last 25 years studying the Bullwhip Effect. This is a dynamic effect in supply chains, and I’ve been applying engineering techniques, computer simulation techniques, mathematical techniques to these problems while living in a business school and helping companies as well. I find it fascinating, and thank you for letting me speak about it today.
Kieran Chandler: Sure, no problem. The idea of the Bullwhip Effect is what we’re going to be looking into in a little bit more detail today. So perhaps you could just start off by giving us maybe a brief overview.
Joannes Vermorel: So, my understanding of the Bullwhip Effect is fundamentally a phenomenon where the fluctuation that you observe in a system, and here we’re looking at supply chain as a system, exceeds the magnitude of the fluctuation that is feeding the input of the system, and typically that would be the demand. And so that would be the sort of phenomenon. And by the way, if you have something that can magnify the fluctuation of the inputs, you also have things that can actually diminish the fluctuation. And that’s typically what can amplify or diminish are things like inventory buffers, for example.
Kieran Chandler: Okay, so Stephen, what are these kind of factors that can influence the so-called Bullwhip Effect? What are they?
Stephen Disney: Well, Hau Lee’s paper in 1997 identified four key sources of the Bullwhip Effect. One is called the demand signal processing, which is about how we interpret demand, forecast it, and then generate replenishment orders. So, there’s a forecasting and replenishment algorithm there. The next one is batching, where we might produce in a minimum order quantity or an economic order quantity, which introduces variability. Another effect is rationing and gaming; you might be short of products or your supplier might be short of products, so you might over order in order to get the products that you actually need, and then you end up cancelling your orders when those products arrive. The final one is price variations. Companies like to manipulate the demand for products, and that can cannibalize future demand for products. The two-for-one offer in the supermarket is a classic example. I buy twice as much toilet paper when it’s half price because it doesn’t go out of date and I have plenty of storage space at home. However, everyday low pricing can help solve that.
Kieran Chandler: Revenue management techniques as well?
Joannes Vermorel: Yeah, the fundamental one, the one that is due to the structure of the system, the lead times, is the forecasting and the replenishment system, the lead times in the demand signal processing cause.
Kieran Chandler: And Stephen mentioned there that demand signal, which kind of manifests itself as forecasting, is the one that seems to get the most attention out of those four factors. Do you think that’s fair? Do you think that’s the right way to go about things?
Joannes Vermorel: I think it depends on the verticals. My own observation is that there are many verticals where what dominates, and yet you have this kind of magnification of variation effect taking place, or the Bullwhip Effect, but the root causes are completely different and exceedingly simple. For example, in retail for fresh food, what drives that are actually stockouts. Because what happens when you have a stockout is that you tend to synchronize the consumption patterns of your clients. So, what we have observed with many food retail companies is that you can observe large fluctuations of demand, and it’s just that you have a population of clients who, when there is a stockout, delay their consumption a little bit. So you end up with an exacerbation effect that synchronizes consumption patterns just driven by the stockouts. Stockouts themselves can actually drive a lot of those large fluctuations by synchronizing the clients themselves. But the bottom line is that I believe that these sort of effects are indeed ubiquitous, but the way they are played out really depends on the industries you’re looking at. The original paper was extensively focusing on FMCGs, and what I’m saying is that it tends to happen in ways that are quite different if you’re looking at companies that are not FMCG companies.
Kieran Chandler: Okay, Stephen, let’s talk about the main conclusions of this paper. It was published over 20 years ago, so what were the main conclusions, and would you say they’re still relevant today?
Stephen Disney: I think the conclusions are very relevant today. They mainly said that Bullwhip was inevitable, especially from the demand signal processing cause. Positive low correlated demand will always generate a Bullwhip effect. And what I see in companies is, typically at an individual product level, production or distribution will be twice as variable as demand, sometimes as much as five times as variable as demand. And this has a consequence on inventory as well, which may be five or ten times more variable than demand. So I see a lot of companies suffering from this effect, and for capital-intensive companies at the bottom of supply chains, this is a big inefficiency. The main conclusion is very relevant, especially in the age of global supply chains, and we have become more global over time with longer lead times. Where I disagree is with the “inevitable” aspect of it. There are things that we can do; we can select more appropriate forecasting methods.
Kieran Chandler: We can tune forecasting methods in the right way, and we can also make those replenishment decisions using ideas and theories from control engineering to help smooth and eliminate the bullwhip effect. We can use information, EPOS data in retail supply chains that can help us. We can also use techniques like vendor managed inventory where your supplier has access to your inventory information and can use that in his decisions. All of these things can help mitigate the bullwhip effect, and in some cases, we can eliminate it completely. Joannes, what are your thoughts? Would you agree with this idea that it’s not so final, and there are ways and means in which you can dampen the impact?
Joannes Vermorel: My perspective is that when we look at the forecasting angle, this perspective, which was late 20th century, is firmly anchored in the point forecast. It’s a forecast that is essentially time series with one data point per year, per day, per week, or per month, and you roll out the forecast forward with a time series perspective and inventory management in mind. What Lokad has done for over a decade is to move towards a probabilistic forecast for all the areas where there is uncertainty, such as demand and lead times. This sort of curse that you had with point forecasts, where either you have a forecast that is always lagging behind, or if you have something more reactive but then you end up with variations that are much greater, those are numerical stability problems that are very much dependent on the fact that we are talking about point forecasts in the first place. When we move to the realm of probabilistic forecasts, most of those problems literally entirely disappear. That would be part one of my answer.
The second part is the way I typically approach supply chain is to say that it’s the mastery of optionality, and part of the game is to cultivate more options. I believe that the sort of vision where you see demand as a monolith and lead times as a monolith is also, to some extent, a little bit dated. First, there is a large amount of substitution. Sometimes you can even engineer the fact that you can leverage this substitution to deliver better service. There might be components that are shared between products that you’re serving, so you don’t naturally have to follow the archetype of, for example, the pharma industry where you
Kieran Chandler: Joannes, what do you think about the idea of keeping options open in terms of transportation and packaging to mitigate potential supply chain issues?
Joannes Vermorel: You need to have the raw active product, but then you can have 150 different packaging options done at the last minute to keep your options open. Nowadays, there are plenty of transportation options available, like air, sea, rail, and road. It’s not that any option is set in stone; it’s a gradient of things that can be more or less expensive. Depending on the situation, you may decide to have an early shipment by aircraft at a much greater cost, just because it’s going to vastly compress your lead times and mitigate a pending stockout. But you’re not going to do that for all your production, just a portion. So, while I agree with Stephen’s conclusion that supply chain issues aren’t inevitable, I would say that the number of ways to mitigate them and make them more profitable for your supply chain has expanded enormously during the last two decades.
Kieran Chandler: Stephen, you mentioned this idea of using control theory, which I see as more of an engineering technique. How can that be applied in this kind of scenario?
Stephen Disney: I have an analogy that I like to use – it’s about taking a shower. In a supply chain, we make a decision, and after a period of time, we receive products either from our production system or from our supplier. There’s a delay between the cause, the decision, and the consequence, the products arriving. Now, imagine we have one of those old-fashioned showers with separate hot and cold taps. To regulate the temperature, I would turn on the hot tap all the way, wait for the hot water to come through the shower and fall on my head, and then use the cold tap to regulate the temperature. If I turn the cold tap too quickly, it’s going to get too cold, and if I turn it back too quickly, it’s going to get too hot. We know that in the shower, we should turn the tap slowly and wait for the water to come through the pipe to match the desired temperature.
The same principle applies to a supply chain. If demand picks up, perhaps because we’ve moved into new markets or our products have become more favorable, we don’t want to chase all of the increase straight away, as we will create oscillations in the supply and demand. We actually want to respond slowly to changes in demand. If we do so and the demand is short-lived, it might drop, and we don’t chase it all the way up or down.
Kieran Chandler: So, Joannes, you mentioned this replenishment algorithm that helps to smooth out variability in production. Can you tell us a bit more about that?
Joannes Vermorel: Yes, absolutely. We get this nice smooth pattern with production orders or replenishment orders flowing through the middle of the peaks and troughs in the demand, and the point in which you do this is in the replenishment algorithms. The forecast is used in the replenishment algorithm in your ERP system, and typically for a high volume product, it will be a variant of something called the water up to policy, and it has two feedback loops in there, an inventory one and a work in progress one. So, we have a target inventory which is our safety stock and our actual inventory could be below or above. Rather than trying to correct all of it in one decision, what we want to do is correct it slowly over time to smooth out the variability being placed on production. It’s the same with WIP, work in progress. If we have a long lead time, there will be a target amount of products in boats in containers being shipped to us, and we have to account for that in exactly the same way as we do for the inventory. But it’s a small change to an algorithm that can have a big impact on the dynamics of supply chains.
Kieran Chandler: Stephen, what are your thoughts on that kind of analogy? It seems quite obvious to follow this idea of almost like a feedback controller. It sounds like it kind of works on the surface.
Stephen Disney: Well, Kieran, I mean, in modern surface, I mean one of the biggest breakthroughs of deep learning actually was a rediscovery of the power of stochastic gradient descent which is exactly that. You nudge a system a little bit touch by touch in the direction where you’re learning. This is what stochastic gradient is about, so it’s like the shower analogy is just a series of small touch you know, hot and cold until you convert. It’s very interesting because two decades ago, people were actually very skeptical about essentially what is now known as local optimization. So, basically, you just follow the gradient and you will get a highly optimized output, and people were thinking, “Oh, if you do that, you’re going to be stuck in sort of local minima and so it doesn’t work.” The reality is that when you’re playing a very high dimensional games, local minimals are not the problem, it’s speed of convergence, and nudging the system with small movement as it is done in the stochastic gradient descent works very, very nicely. So, that would be a part of my reflection. Then, there is another thing which is when we are saying of what we are trying to optimize, again, I would say it very much depends on the sort of verticals that you’re looking at, because, for example, let’s consider hard luxury. Let’s say you’re producing your master watchmaker and you produce very expensive watches, and let’s say the extreme would be women’s watches. So, what you have is essentially a piece of jewelry which is made of precious metals, 100% recyclable. You can recycle 100% of the value. You have like precious metals, gemstones, and then you have a movement which is kind of standardized. So, what are your constraints? I mean, you can and literally, if you don’t have something to show in the store, then people don’t buy, so you have it.
Kieran Chandler: of your interest to produce really a lot and it’s a market, hard luxury is very much driven by novelty so you need to produce a lot. And what happens if you don’t produce?
Joannes Vermorel: Well, what you don’t produce, you just bring back the expensive watches to a store, you dismount all the precious gemstones, you recycle the metal, and you put the movements into new watches and then you send them back. Then you realize that maybe the assembly is something like five percent of the cost of a watch and that it’s 80% gross margin. So, you see, when you’re in this sort of situation, you have very strong asymmetries at play. Obviously, it varies. It’s not the same thing for a highly pressured FMCG that really operates on super tight margins. My point is that gradient-based optimization really works, but then you really need to think in terms of the asymmetry of the economic drivers to know what are the areas that represent the sweet spot in terms of equilibrium. And from one vertical to another, what would be considered as insanely wasteful in one industry might actually be considered as very reasonable in another industry.
Kieran Chandler: Okay, Stephen, let’s talk a little bit maybe about the application of these techniques in the real world. I think one of the things the paper was very good at is it highlighted some of these issues, but it wasn’t very prescriptive in how to deal with them. So, what is the kind of advice that you give to companies that you perhaps work with?
Stephen Disney: Building on Joannes’s ideas, I think the first step is to understand the needs of your supply chain. So, value stream mapping is an important first step. You want to understand the process that is used to produce a product. You want to understand what are the strategic needs of that process. Is it a supply chain where you’re focusing only on reducing inventory, or are capacity costs significant? If you’re only focusing on inventory and bullwhip has no consequence, then focus on minimizing your inventory costs, and that’s fine. In capital-intensive industries, it’s probably a balance between inventory costs of finished goods and raw materials and the efficient use of your production facilities and the capital that you have tied up there. So, understanding your supply chain, mapping it – I like to use value stream maps to do that – and then overlaying on that, time series of what’s the time series of demand, what’s the time series of forecasts, what are the time series of production targets and production completions, what’s the time series of the inventory levels, the finished goods, the raw materials. And then going back to your supplier, what do the forecasts look like, the order call-offs that you give to your supplier? Do they match their deliveries, and are you able to give reliable future guidance to your supplier about what is needed? That will give you an understanding of where the variability in the system is being generated and what are the consequences of that variability because it’s not always bad. So, when you’ve understood your strategic needs of your supply chain, you can then start thinking about how you’re forecasting. Often companies will manually adjust their forecasts, and we need to think about whether that actually adds value compared to an algorithmic forecast.
Kieran Chandler: Are you using the right forecasting algorithms? Are they tuned correctly for the needs of your business? An inventory-focused supply chain will have a different forecasting need than a capital-intensive forecasting company. We must remember, we’re not creating forecasts to show people how well we can predict the future, but we’re creating forecasts to make a business decision about how much to order from our suppliers and how much to produce. So, the algorithms that use those forecasts, are they set up in the right place? Do they have proportional feedback controllers, the speeds of the taps? Are they set up correctly?
Stephen Disney: That’s work that can be done in your IT system, your ERP system, or your spreadsheets that you use to plan production and source from suppliers. And then there’s good old-fashioned engineering work to be done. Is the production system able to produce to the desired plan? Are your machines reliable? Do you hit the production targets, or do you sometimes overproduce or underproduce? Do you produce good quality products? It’s a mix of forecasting, computer science, control engineering, and good old-fashioned production engineering to get the bullwhip effect to a level that is appropriate for your supply chain.
Kieran Chandler: Brilliant. And Joannes, how relevant would you say the bullwhip effect is today in the present day? Would you say that COVID was a very good example of the bullwhip effect in action?
Joannes Vermorel: I mean, COVID was a super large-scale disruption for everybody. I don’t think it was exactly a manifestation of the bullwhip effect. I would say if there’s one thing that characterized COVID, it was a sort of fat-tail event. It was a reminder that distributions in supply chains, and by distributions, I mean statistical distributions, are not normal; they are fat-tailed. So, you have these extreme events that are not as unlikely as they appear when looking at normal distributions.
I suspect that due to the erraticity in lead times, there will be all sorts of problems as a result of this pandemic that will take the form of bullwhips. To some extent, I think that’s what we’re seeing in Asia right now for electronics. But I don’t believe it’s going to be dominant. Nowadays, I think what is very interesting is that if you take an approach where you look at all possible futures and then look at all possible decisions and cross the two, you can have a very granular response that was not technically feasible two decades ago. You can really quantify how far you are willing to go into having variations in your system that exceed the variation in demand because, usually, in supply chains, things have non-linear costs.
Kieran Chandler: If you want to produce twice as much during the same time frame, it may not cost twice as much; it may cost five times as much, just because people need to work overtime, machines will have to run at a level where aggressive maintenance is needed, and so on. The question is, can you eliminate these problems or at least put them under control from a financial perspective, where you have much more control over the financial outcome for your company?
Joannes Vermorel: Nowadays, I don’t think you can eliminate those problems, but you can largely put them under control from a financial perspective, giving you much more control over the financial outcome for your company.
Kieran Chandler: Stephen, we’ll leave the final word to you. I know you do a lot of research in the field of statistical techniques and applying them to operations management. What are you researching at the minute, and what do you think will be of interest over the next coming years?
Stephen Disney: I’ve spent a lot of time recently looking at dual sourcing supply chains. These are supply chains where we source the majority of our product from a low-cost country that might be far away, but we supplement that with a small local factory. The small local factory will have a shorter lead time and might be more expensive per unit to produce locally, but because we can satisfy most of the demand from the long lead time, low-cost supply chain, the unit cost on average is quite small. The small factory can flex its volume very quickly to accommodate the variability in demand, so we can keep our inventories under very tight control with the small factory while satisfying most of the demand from the low-cost products from the distant factory.
Looking at how we exploit these ideas is interesting. I think it’s good from an environmental point of view because the net distribution of products around the world is going to go down. It’s an interesting way to bring back manufacturing into the expensive Western countries and hopefully will make our supply chains more robust to disruptions.
Building on the last question, to me, the supply chain is a system with a natural frequency, just like a bridge will vibrate in the wind at a certain frequency. The supply chain has a natural frequency, and we’ve just given it a big COVID kick. The supply chain is going to oscillate at its natural frequency for a few years before that oscillation dies down. We’re going to see demand go up, then drop, and then pick back up again. Global supply chains with long lead times will take longer to dissipate, while short lead time supply chains will recover a lot quicker.
Kieran Chandler: Brilliant. Well, thank you both for your time. That’s everything for this week. Thanks very much for tuning in, and we’ll see you again in the next episode. Thanks for watching.