00:00:07 Introduction and discussion on paradigm shifts in supply chain planning.
00:00:51 Professor David Simchi-Levi’s background, academic work, and companies he started.
00:02:43 New course on demand planning and analytics; focus on digitalization, analytics, and automation.
00:04:22 Balancing supply chain efficiency and resiliency, and the challenges in measuring resiliency.
00:07:08 Risk assessment in supply chains and the importance of machine-driven calculations over intuition.
00:09:47 Predicting supply chain state weeks in advance.
00:10:20 MIT team’s work on supply chain resiliency and pandemic prediction.
00:12:15 KPIs vs. KPPs and the importance of predictive data.
00:15:35 Embracing uncertainty and different likelihoods in predictions.
00:18:10 Time to survive and identifying hidden risks and cost-saving opportunities.
00:19:32 Importance of explainable machine learning in decision-making.
00:20:56 The role of supply chain scientists in crafting explainable metrics.
00:24:17 Challenges faced by companies when dealing with black box forecasting.
00:25:51 Laws of physics in supply chain and their significance.
00:27:34 Examples of laws of physics in supply chain management.
00:29:16 Discussing scientific debate and the importance of mathematical foundations in building algorithms and frameworks.
00:30:57 Importance of flexibility in supply chain design and its impact on service level, inventory, and response time.
00:32:29 Addressing the economic trade-off in flexibility and the need for foundational theories in supply chain optimization.
00:35:44 Focusing on frameworks rather than single solutions to accommodate the changing nature of supply chains.
00:37:18 Introducing four frameworks for supply chain digitization, starting with a unified view of demand.
00:38:14 Supply chain segmentation for effective strategies.
00:38:53 Focus on S&OP and data analytics for planning.
00:39:37 Jonas discusses the complexity of demand and enterprise systems.
00:43:01 Challenges in extracting and understanding demand data.
00:45:36 The shift from forecast and plan to numerical artifacts. {< timer “00:47:14” >}} Discussion on decision-makers relying on outcomes rather than prediction or plans.
00:48:22 Emerging algorithms in supply chain learning and optimization, and negative trends.
00:49:33 Misguided concept of moving manufacturing closer to market demand for resiliency.
00:50:09 Integrating machine learning and optimization, offline and online learning.
00:53:22 Challenges of multi-layered enterprise software and new algorithmic paradigms to simplify processes.
00:56:37 Discussing the IT complexity in supply chain management.
00:57:57 The need for tools to handle complex concepts like resiliency and risk management.
00:58:43 The advantages and limitations of Microsoft Excel in supply chain decision-making.
00:59:30 The reluctance to invest in supply chain digitization before the pandemic.
01:00:20 The opportunity to improve business performance with moderate investment in supply chain digitization.
Summary
In this interview, Nicole Zint hosts a discussion with Joannes Vermorel, founder of Lokad, and Prof. David Simchi-Levi, an MIT professor, about supply chain optimization and technology’s role in decision-making. They emphasize the importance of balancing efficiency with resiliency, using data, analytics, and automation to manage risks. Key Performance Indicators (KPIs) and Key Performance Predictors (KPPs) are introduced as essential concepts for proactive decision-making. The experts also discuss the importance of explainability in AI, “laws of physics” in supply chain management, flexibility, and the economic trade-offs between flexibility and modeling. The conversation highlights the need for companies to digitize their businesses to address future challenges and opportunities.
Extended Summary
In this interview, host Nicole Zint speaks with Joannes Vermorel, founder of Lokad, and Prof. David Simchi-Levi, an MIT professor and author of over 300 publications, about supply chain optimization and the role of technology in decision making. The discussion explores the paradigms of learning and adapting in supply chain management, with a focus on the integration of digitalization, analytics, and automation.
Prof. Simchi-Levi highlights that the current state of supply chains is vastly different from a decade ago, and companies must balance efficiency with resiliency. He notes that while efficiency is easy to measure through cost-cutting strategies, resiliency is more challenging to quantify. The integration of data, analytics, and automation plays a crucial role in addressing these challenges by helping businesses identify and mitigate hidden risks.
Joannes Vermorel agrees that the trade-off between resiliency and efficiency is essential, emphasizing the probabilistic nature of risk measurement. He argues that survivor bias is prevalent in market evaluations, making it difficult to assess the full extent of risks faced by companies. Vermorel underscores the importance of adopting a machine-driven, calculation-based approach to risk management, as opposed to relying solely on human intuition.
Prof. Simchi-Levi introduces the concepts of Key Performance Indicators (KPIs) and Key Performance Predictors (KPPs). KPIs focus on the current state of a supply chain, while KPPs aim to predict the state of the supply chain in the future. The professor emphasizes the importance of using data and analytics to complement KPIs with KPPs, allowing companies to take corrective action before problems emerge.
Vermorel supports the distinction between KPIs and KPPs, highlighting that many supply chain directors may not realize the predictive component in their KPIs. He explains that real-world supply chain systems are often messy, with inputs that are not entirely reliable, especially when predicting the future. Vermorel advocates for embracing uncertainty when dealing with future predictions, as even a well-informed guess can provide valuable insights for decision-making.
Prof. Simchi-Levi uses a sports analogy to illustrate the importance of KPPs, referring to hockey player Wayne Gretzky’s famous quote, “I don’t skate to where the puck is, I skate to where the puck is going to be.” In supply chain management, this means making decisions today to address potential challenges in the future, ensuring the adaptability and resilience of the supply chain.
The discussion revolves around supply chain optimization, embracing uncertainty, scenario analysis, and the use of machine learning and AI in supply chain decision-making.
Prof. Simchi-Levi emphasizes the importance of scenario analysis in supply chain management. By generating multiple scenarios based on various factors such as changes in demand or disruptions, companies can identify hidden risks and cost-saving opportunities. However, he acknowledges the limitations of this approach, citing the unpredictability of real-world events. To overcome these limitations, he suggests using criteria and tools that are independent of specific scenarios or supplier information. He gives examples of concepts he has developed, such as “time to recover,” “performance impact,” and “time to survive.”
Both Simchi-Levi and Vermorel agree on the importance of explainability in machine learning and AI for supply chain management. They argue that human supply chain planners will not trust recommendations from a machine if they cannot understand how the machine arrived at its conclusions. Vermorel suggests that a human “supply chain scientist” should work alongside AI algorithms to craft the predictive optimization logic and its explanatory factors, while acknowledging the human expertise required for effective decision-making.
Prof. Simchi-Levi introduces the concept of “laws of physics” in the context of supply chain management. These laws are general relationships between various supply chain factors that are universally applicable, regardless of the industry or specific supply chain. He provides examples of such relationships, such as the link between inventory safety stock, service level, and variability. These laws can help companies manage their supply chains better by understanding the underlying principles governing their operations.
Vermorel supports the idea of universal laws in supply chain management, citing Zipf’s law as an example. He explains that this law can be observed in various aspects of supply chain data, such as product distribution and supplier size. The knowledge of these laws can be instrumental in building effective tools, algorithms, and frameworks for supply chain optimization.
The discussion also touches upon the importance of flexibility in supply chain management. Prof. Simchi-Levi explains that while flexibility is essential, it does not come free. Companies must understand the right amount of flexibility they need, where to invest in it, and its potential benefits. Using the laws of physics, supply chain managers can design their supply chains for flexibility and quantify its impact on various aspects of the supply chain, such as service level, inventory, and response time.
They talk about the importance of finding a balance and the economic trade-off between flexibility and modeling. The focus is on establishing a core of reliable theories in supply chain management to build upon. They also discuss the need for a long-standing strategy to avoid the constant need for change. Professor Simchi-Levi mentions four frameworks for supply chain digitization, including a unified view of demand, supply chain segmentation, and effective planning. Joannes Vermorel emphasizes the complexity of data and the importance of tools to extract and process it. He also describes a method that focuses on day-to-day execution with no plans or forecasts. They conclude with an example of fashion retailing and how prediction is used to manage the supply chain.
They discussed the emerging trends and negative trends in supply chain optimization were explored. The group discussed how the new normal of disruption and volatility is impacting the supply chain industry, and how companies need to rethink their management strategies. The integration of machine learning and optimization was also discussed as an opportunity for better decision-making, with offline and online learning being key components. However, the complexity of modern enterprise software and the need for better tools to drive decision-making were seen as negative trends hindering supply chain initiatives. The group agreed that companies need to take the opportunity to digitize their businesses to address future challenges and opportunities.
Full Transcript
Nicole Zint: In today’s episode, we are discussing these paradigms for learning and optimizing supply chains. We are honored to be joined by Professor David Simchi-Levi. Today, we’re going to be discussing his work and his over 300 publications. Professor, as always, we would like to kick off by our guests introducing themselves. Thank you.
Prof. David Simchi-Levi: Hi, Nicole. Hi, Joannes. Great to be here. I’m David Simchi-Levi, and I’m on the faculty at MIT. I’ve been an academic for a long time, just at MIT over the last 21 years. But in parallel to being an academic, I started a few companies. The first company was in supply chain analytics, a company that became part of IBM technology infrastructure in 2009. At that time, we had about 350 clients using our technology for multi-echelon inventory optimization, supply chain network design, and related topics. In 2011, I started another company around business analytics. The focus was not only on supply chain but beyond just supply chain. This company became part of Essential Technology in 2016. Then, I started a cloud technology company in 2014 that became part of Accenture in 2018. Currently, I’m company-free and focusing on my research at MIT. At MIT, I lead the MIT Data Science Lab. The Data Science Lab is a partnership between MIT and about 20-25 companies focusing on addressing some of the most important challenging problems they have by bringing together data, models, and analytics. Hopefully, we’ll have an opportunity to talk about some of the exciting work and opportunities that we see today in the market.
Nicole Zint: That’s a very impressive background, I must say, Professor. In addition to all that, you also recently launched a course, Demand Planning and Analytics, where you mentioned these three emerging technologies: digitalization, analytics, and automation. Why are those so popular right now?
Prof. David Simchi-Levi: It’s interesting that you mentioned the new class that we just launched around demand and supply chain analytics. In this class, as you pointed out, we are focused on the integration of data, analytics, and automation. It’s really about the integration of these capabilities that companies are able to address some of the most challenging areas they face in their business. Think about supply chains with long lead times, the significant increase in logistics costs because of changes in oil prices, and supply chain disruptions that we have seen over the last three to four years. From the U.S-China trade war to COVID, the Ukraine war, all the way to climate change, all of these have required companies to rethink their supply chain strategy today.
Nicole Zint: The normal is completely different than what we saw ten years ago. How do you manage supply chain effectively today? It’s different than what companies did five to ten years ago, and I will summarize with one example. Until about 2020, there was a lot of focus in the industry on supply chain efficiency, from lean to outsourcing to offshoring. Companies focused on cutting costs dramatically in their supply chain. But what they have observed over the last three years is the need to balance supply chain efficiency and resiliency. Efficiency is easy to measure; you focus on cost. Resiliency is not easy to measure. How do you measure resiliency? How do you identify hidden risk? This is all about the technology trends that I just mentioned: digitization, analytics, and automation. So Joannes, what do you think about this compromise between resilience and efficiency that the professor just mentioned?
Joannes Vermorel: It is indeed a trade-off in the sense of an economic trade-off because resilience is typically not free. It’s about building up your options that you need to maintain and establish. The interesting thing is I completely agree with Professor Simchi-Levi in the sense that it’s very difficult to measure because fundamentally, you’re talking about a probabilistic perspective of the future. You’re looking at things that may or may not happen. For example, if you invest in having a second line of suppliers that happen to be nearshore but don’t use them this year, you see the cost, but you don’t see the very existence of the option that would save you if you were to need it.
What’s interesting is that you have survivor bias all over the place. The only companies that you see are the ones that are still alive. Those that made a terminal mistake and disappeared are not there anymore, so you don’t see them. When you look around and do a survey, you always see, on average, people taking too much risk compared to what they should. The reason for this bias is that you consistently have people who took too much risk and exited the market, but when you do a survey, they are not part of the survey anymore because you only survey active companies.
The trick, or more than a trick, one of the first paradigms of risk measurement, is to have an assessment of potentials that are, most of the time, unrealized but nevertheless very real. If you roll a dice and have only a 3% chance to hit a critical event that would terminate your company, and you do that every single year, over the course of half a century, you’re almost certain that your company will disappear due to those long-term events.
Nicole Zint: Supply chains have typically been built over dedicated, I mean, very large companies. There are some companies that are very large today, like, let’s say, Apple, but even Apple is not exactly like a brand new startup. It took decades to establish and become what they are. So even companies that did grow fantastically fast, for them to reach the kind of maturity of the supply chain, we are still talking about a multi-decade process. So, it is slow, and assessing risk when you’re thinking in terms of multi-decades period, you have to look at things that are happening very rarely from, I would say, the human perspective. That’s also why the sort of machine-driven perspective, calculation-driven as opposed to sheer intuition, becomes so important. Joannes, can you share your thoughts on this?
Joannes Vermorel: I believe that humans are very good at perceiving things at a human scale, and supply chains tend to diverge from that in complexity. There are just too many things, but also in terms of time span, we really think in terms of things that could happen once or a quarter in a century. And yet, if you operate a large supply chain, this is the sort of risk you should think of.
David Simchi-Levi: Let me demonstrate what Joannes was emphasizing with a key observation. Companies typically focus on KPIs. They ask, “What is the performance of my supply chain right now?” If the service level is low, they might make a change, like adding more inventory. If transportation costs are high, they might make a change to cut costs. This discussion about resiliency, this discussion about using data and analytics, is also about complementing KPIs with what I call KPPs - key performance predictors. Everything might be looking good in a supply chain now, but we’d like to predict what will be the state of the supply chain six or seven weeks from now. Because if we can do that and see a potential problem, we can solve the problem today before it hits the supply chain seven weeks from now.
And you may think, “Hey, is it possible?” In fact, the pandemic showed us that we can do this very effectively. Let me illustrate this with a story. My team at MIT’s Data Science Lab worked on supply chain resiliency way before the pandemic. We developed a new way to measure the resiliency of a supply chain to identify hidden risks. We implemented it at a number of companies. The first company was Ford Motor Company, then other companies followed suit, but not many. Everything changed at the beginning of the pandemic.
In February of 2020, remember this time period, the pandemic was hitting China. It was not in Europe, it was not in the US, it was only hitting China. I wrote a very short paper using the model and the data that I had, using the model for supply chain resiliency, and I wrote a short paper that said in mid-March, so this is six weeks later, we will see a disruption of supply chains both in North America and in Europe. And this is exactly what happened. So the ability to use data, and we have both real-time data that come from internal data that companies have, plus external data, allows companies to complement their KPIs, key performance indicators, with KPPs, key performance predictors, which is the state of the supply chain right now.
Nicole Zint: With key performance predictor, what will be the state of my supply chain six weeks from now or eight weeks from now, and take corrective action today before the problem hits the supply chain? That’s why Joannes was emphasizing, for one reason, the importance of using data and analytics machines to make a big impact on supply chain performance.
Prof. David Simchi-Levi: I really like this distinction between KPI and KPP. By the way, I believe that most supply chain directors in most companies don’t realize that actually a vast majority of what they call KPIs do include, one way or another, a prediction component. For example, if you say that you have KPIs about service levels, most of the service levels actually include a demand forecast. When you say that you have this level of service or this amount of service level, the reality is that for most SKUs, demand is very sparse. You do not have an SKU where you say, “I have 90% service level.” It’s either present or absent. What you do have is, eventually, a step-by-step analysis or a predictive model of some kind that would give you this sort of estimate of what your service level is for your SKU, but it is also an estimate. And even then, there are frequently situations where, for example, the amount of stock they have actually depends on stuff where they have ETAs that are not completely guaranteed.
Joannes Vermorel: So the binary, I think, having the two concepts is very interesting, especially to realize how in real-world supply chain systems that are very messy, where there is a lot of input that is internal but not completely reliable, especially when you touch the future, you have a gray area that is very blurry between the two. And many people, I think, one of the problems is first to have too much trust in believing that it is like a neutral objective, I would say, observation of the past for many indicators. It’s not. And then there is another aspect, which is as soon as you want to deal with risk, it means that you can’t have this naive one-future perspective anymore. You need to think that, when you predict that there will be a disruption that would come, you would most likely, if you were very careful, say something like, “I am 80% confident that there will be a disruption starting from eight weeks at the soonest and twenty weeks at the latest,” etcetera. But fundamentally, it is to embrace the fact that as soon as you touch the future, you have things that are likely but not certain. And yet, it has value. It’s not because there is a degree of uncertainty that suddenly you should say it doesn’t exist. I mean, having a very good guess, even if it’s just a guess, is already something that has a lot of value, and you should act upon it, even if it’s just probabilities.
Prof. David Simchi-Levi: Let me, Nicole, if you allow me, illustrate why companies need to think about KPI and KPP with an example from sports. I never played hockey, but I like hockey. And if you think about one of the best hockey players in North America, Gretzky, he used to say, “I don’t skate to where the puck is; I skate to where the puck is going to be.” This is really what KPP is all about.
Nicole Zint: I want to make a decision today to address where my supply chain is going to be six weeks from now or seven weeks from now because if I can do this today, I can reserve capacity, cut my cost, and respond effectively to a potential disruption that my system suggests I will see in the near future. So, it’s really interesting that we can predict or supposedly can predict something that can happen in our supply chain, say six, seven, or eight weeks from now. However, at Lokad, we try to embrace this uncertainty and rather that we can’t really predict what’s going to happen exactly, but we want to have essentially an overview over the likelihood of various futures happening. So, professor, when you say that you know you predict something six or seven weeks from now, how exactly can you know? Are you then just sort of focusing on one future scenario that the model outputs, or are you rather embracing these different likelihoods?
Prof. David Simchi-Levi: We are using a combination of approaches. There is no one-size-fits-all strategy, and let me highlight this as I think it ties together with what Joannes and you were focusing on. One approach is scenario analysis. We generate multiple scenarios; the scenarios may be associated with a change in demand, a scenario may be associated with a disruption at a specific supplier, or a specific region. And using this, we are trying to identify hidden risks in the supply chain. But there is a limit to our ability to generate scenarios, and to illustrate this, just think about what has happened in the last three months relative to what happened two years ago. Who would have predicted what we see now in Eastern Europe, right? So, scenario analysis is very important, it’s part of what we do, but we also need criteria and tools that are independent of a specific scenario, for example, independent of information that we may get from specific suppliers. These types of tools still exist and are available for companies today to use. Let me illustrate this with one example. I developed a few concepts around supply chain resiliency: one is Time to Recover and Performance Impact; these are all scenario-dependent. But I also developed Time to Survive. What is Time to Survive? Time to Survive is completely independent of a scenario. You look at the entire supply chain end-to-end, you have a supply chain mapping, you know where your inventory and how much inventory you have, and now you remove a facility from the supply chain and you ask, without that facility, how long can I make supply with demand? This is not scenario-dependent, right? This is given the disruption, and I will tell you how long I can manage the supply. This allows me to identify risks in the supply chain but also cost-saving opportunities. When we implemented this in multiple companies, we realized that sometimes companies throw a lot of inventory in the wrong location for their own product. This allows you to identify hidden risks and allows you to identify savings opportunities as well. So, it’s not one approach that gives us a good understanding of supply chain resiliency; it’s multiple approaches.
Nicole Zint: The last element that I will add is that we focus a lot on using machines, machine learning, and optimization to make better decisions. But nobody is going to follow a recommendation from a machine if the machine cannot explain itself. So, on top of the ability to generate a forecast or a recommended decision, we need an explanation of why this forecast shows that demand for product A will grow significantly in the Midwest but will not be successful on the West Coast. Explainability of what comes out of the machine is a critical part of this decision-making process.
Joannes Vermorel: Jumping on your comments about the need for explainability of the models, my own casual observation is that numerical models, even semi-trivial ones like linear regression with a couple of coefficients, are very much opaque by default. This is kind of a given as soon as you have digits. Computers are so much better at doing calculations than people that it doesn’t take many numbers to have something that is completely opaque to the average non-genius human observer.
And the typical approach is that a number that makes sense in terms of explanation tends to be incredibly context-dependent. It’s very tempting to just produce a wall of metrics with millions of numbers that you can extract from projecting your data in all sorts of directions. It’s very easy to do with modern computers, but your supply chain planners have just a limited amount of time to do that.
Thus, the approach that Lokad has is working on a process where, at an algorithmic level, it will be relatively straightforward for what we call the supply chain scientist (which is like a data scientist specialized in supply chain) to craft both the predictive optimization logic and its explaining factors. But there is a catch: I’m not expecting the AI or any fancy machine learning recipes to be able to do this work. I am more adopting an approach that is paradigmatic, where I say I have classes of algorithms where I know that a supply chain scientist, with their very human intelligence working with that, can do the extra mile of doing what we call white boxing.
This allows them to craft the sort of metrics that will make sense to the supply chain management at large so that they can understand what is going on. But there is a very human ingredient in that, which is to have someone essentially create the numbers so that you can select a few KPIs. Your numbers, not just KPIs, can be TPP according to your definition, but they are very carefully selected. The only magic trick is to have an algorithmic method that lends itself very nicely to this sort of in-depth decomposition of what is going on.
Prof. David Simchi-Levi: Correct, and I can highlight why this is so important with an example of implementing what I call a unified view.
Nicole Zint: Of demand in a very large CPG, when you implement this, you start getting phone calls, typically from the finance group. The phone calls are one of three types. The first is saying, “Hey, data scientists, hey, Tim, we don’t understand why your focus suggests that this product or for a family will grow incredibly well in one region but not do well in another region,” right? That’s part of the explainability.
David Simchi-Levi: The second part is even more challenging. You get another phone call, and the finance people are saying, “We don’t understand, you just gave us a forecast because we give a forecast every week for the next 80 weeks. The forecast you gave us today is different than the forecast you gave us four weeks ago. What is going on? The world has not changed.”
David Simchi-Levi: The third is, “Hey, a month ago you gave us a forecast about today’s demand, but it’s quite different. If we cannot address these three challenges, nobody is going to trust a black box that spits out forecasts weeks in and weeks out for every product. That’s why it’s so important to recognize that what you are doing in your company, what my team at the MIT Data Science Lab is doing, is not just science. To be effective, it requires combining science and art. Science is the machine and the data in the analytics; art is the insight, the intuition, the experience that people, in this case in supply chain, have. It’s a combination of the two. But if we cannot talk to the machine to understand what the machine is saying, it will be hard for humans to follow recommendations from the machine.”
Nicole Zint: So, Professor, before I get to the main questions about which emerging paradigms we have, I just want to ask you, you also mentioned these “laws of physics” in your course, you know, applicable to supply chain practitioners and companies. But what do you mean by “laws of physics” in the supply chain perspective?
David Simchi-Levi: For me, “laws of physics” are general relationships between information, capacity, service level, inventory relationships that are always true, independent of whether you have a regional supply chain or a global supply chain, independent of whether you focus on high-tech, CPG, or pharmaceutical.
Nicole Zint: Could you give an example of one of those relationships that you just mentioned?
David Simchi-Levi: A relationship between inventory, safety stock, service level, and variability. We know how to quantify the relationship between the three: safety stock or inventory, variability, and service level. Once you understand the relationship between the three, you can realize how to manage your supply chain better.
David Simchi-Levi: Another example is the relationship between the level of information and how much volatility I will have in the supply chain. Once you understand that, you can realize how much visibility and information sharing can allow us to reduce volatility. And what is the relationship between volatility and lost sales? If you understand that, you realize, “Oh, I need to reduce…”
Nicole Zint: How can I manage volatility in order to increase my service level and reduce my lost sales?
Prof. David Simchi-Levi: One way to reduce volatility is through information sharing. In my book, which had its fourth edition published in September of last year, I discuss about 40 or 50 types of “laws of physics” that allow companies to identify opportunities in their business. These are global relationships, and that’s why I call them laws of physics.
Think about what we learned in high school or college physics classes. The idea is that there are some fundamental relationships in business that define a sort of quantitative measurement of different parts. You have situations where you can project equations that will be true, like the four Maxwell equations in physics, all the time. It’s not dependent on a particular situation.
Joannes Vermorel: So, essentially, supply chains are not entirely about poetry. There are these fundamental relationships in business that define a sort of quantitative measurement of different parts. For example, in my series of lectures, I point out that virtually every single distribution observed in supply chain is a Zipf’s law. From the highest volume products to the long tail, you will get a Zipf’s law. The same goes for suppliers, from the largest to the smallest, you will get a Zipf’s law, and so on.
This theory can be challenged in a scientific sense, like questioning whether it is the best theory to explain the world or if there are situations that contradict the general theory. However, it’s not up for debate in the sense that you can simply choose not to believe it based on your specific industry.
The interesting thing is that when you have these mathematical foundations for classes of phenomena, it is incredibly powerful for building tools, algorithms, and frameworks. At Lokad, we use this extensively. For example, the Zipf distribution has a consequence that you can compress supply chain data tremendously, as the majority of your lines are guaranteed to be either zeros or ones, making them eligible for compression.
Another example is that you can actually beat the quicksort algorithm in terms of sorting. You can be faster than the theoretical optimum just because of the low cardinality of what you have to solve. There are plenty of things that are important in terms of software design and also the design of the supply chain theories built on top of these mathematical foundations.
Nicole Zint: So, we’re here today with Joannes Vermorel and Professor David Simchi-Levi to talk about supply chain optimization. Joannes, you often speak about the importance of flexibility in supply chains. Can you elaborate on that a bit?
Joannes Vermorel: That brings this to life if you allow me. Everybody understands the concepts, the idea of flexibility. Everybody understands that more flexibility is better than less, but flexibility does not come free. How much flexibility do I need? Where should I invest in flexibility, and what are the potential benefits of flexibility? These are key questions to answer. But on top of this, the question is how do I define flexibility? Once you have a precise definition of flexibility, for example, the ability to respond to change, and change can come in many different ways, changing demand volume, demand mix, disruption, we know exactly how to design a supply chain for flexibility. These come from low physics, and once you know that, then you can quantify what is the impact on my service level? What is the impact on my inventory? What is the impact on my response time? And this is taken advantage of by companies in the automotive industry, in the consumer package industry, using laws of physics to rethink the degree, the level of flexibility in the supply chain. The same is true for redundancy, the same is true for resiliency. This is why laws of physics are so important.
Nicole Zint: It’s quite interesting that you mentioned flexibility because we often see when companies have to make decisions on whether to keep a particular product in their central warehouse or more local warehouses, you have this essentially this balance or this compromise between being more flexible when it’s in the central DC or you have a better service to your customers because they get the product quicker.
David Simchi-Levi: So yes, indeed, it is about finding this sort of perfect balance. But I think the point most specifically that Joannes is making is that it is not just about finding the balance, about finding the balance is kind of a given. What he’s pointing out is that there is a trade-off, an economic trade-off that is ever present with flexibility, and that it can be modeled, and that it will be relied upon to do optimization in all situations. You see, this is what he’s saying, I believe that’s this kind of nickname of law of physics. This is the law of, I would say, physical supply chains. You know that is like the shorthand for that. There are foundations that can be, that have been established. They are not naturally complete; they are not naturally definitive, but they are there. And it is important to approach, you know, those supply chains here in this case with a mindset where everything is not, I would say, up for debate. You see, that’s the point of this sort of approach of physics. When people say we have the equation for electromagnetism, you know that we’re essentially what we call the four Maxwell equations, so we’re not done by Maxwell, but whatever, the four equations of electromagnetism. You can try to disprove them, but meanwhile, everybody is going to do electromagnetism, saying this is what I consider as true all the time. This is not an option. People don’t say that it’s naturally impossible to find a situation where they would be at fault. This is not science. It’s about discovering the flow in your model, and we find further. But the point is that I think that’s very interesting to have to establish, I would say, a core.
Nicole Zint: So Joannes, could you tell us about the theories that you rely on in order to build a quantitative approach that works for supply chain optimization?
Joannes Vermorel: There are two theories in the scientific sense that can be relied upon so that you can build on top. You know, as opposed to only having opinions and debates. Because the problem is that if you don’t have this score, it becomes very difficult to have a quantitative approach that is not accidental. You want to have a method that has the potential to work all the time in all companies. That’s the pinnacle of generality. But if you have very solid foundations, you can have things that get closer and closer to that. I think that’s where the true interest lies.
Nicole Zint: So it’s interesting that you’re saying finding this solution that can work for many different problems, not just one. And in fact, Professor, that’s something we’ve also seen in your publications. That you often talk about not just a single algorithm or a single solution, but more of a framework that can be applicable to a whole different sorts of problems. So why exactly have you focused on that rather than just handing out single modules?
Prof. David Simchi-Levi: Well, it’s probably very obvious, but at the same time, we see that’s not quite often that people do focus on the framework. But it’s always finding this one solution. But then when the supply chain is changing, and you mentioned, you know, if the world is changing all the time, then we’re stuck. So we need to redo the problem again and again and again. And in fact, at Lokad, we focus on exactly that. Finding the sort of problem-solving approach rather than that one single solution that’s only applicable right now.
Nicole Zint: Professor, could you also answer the question about why you focus on frameworks?
Prof. David Simchi-Levi: In your description of the focus of my research, a lot of the focus is to make sure that we have a long-standing strategy for the supply chain. So that we don’t need to change our strategy every week or every day with a recent disruption or change in demand. And maybe I will highlight this with the work that I’ve done around supply chain digitization. What are the frameworks that we have identified in supply chain digitization that allow companies to achieve most of the benefits of full digitization without the four or five years investment in digital supply chain? And I will highlight four frameworks. The first, I mentioned earlier, is the unified view of demand, replacing consensus forecast. Consensus forecast has been used by executives and industry for many, many years. Finance will come as its own forecast. Operation will have its own forecast. Sales will have a different forecast. And then they will get together in a consensus meeting to agree on a compromise. Not clear that this compromise correctly represents reality. What you want to do, the framework I focus on, is to agree on the data. Once I have the data, I would like the analytics and machine to generate a forecast that can be used by the different functional areas. That’s the first one. The second one is a…
Nicole Zint: Could you tell us about the framework that you use in supply chain optimization?
Prof. David Simchi-Levi: Sure, so the framework that we use is based on the idea that a one-size-fits-all strategy is not appropriate for most companies. If you look at what most companies are doing, they have one strategy across all channels, across all markets, and across all products. What we emphasize is supply chain segmentation, segmenting products, segmenting markets, and segmenting channels. This allows companies to fine-tune the supply chain strategy for each cluster, for each group, and as a result, be more responsive depending on the characteristics of each segment. The third element in this framework is focusing on SNOP (Sales and Operations Planning) that uses data and analytics to help companies identify an effective plan. The last one is recognizing that as effective as the plan is, there are always deviations from the plan such as supply disruption and demand changes. If I can identify those disruptions and deviations early on, I can respond to them very effectively. This is part of the KPP (Key Performance Parameters) and control tower, which I’m sure your company is focusing on with many of your clients.
Nicole Zint: Joannes, what do you think about what Professor Simchi-Levi is saying, especially about this SNOP process, which Lokad also has?
Joannes Vermorel: Yes, um, but it’s not, I would say, a different approach. Due to our focus and our origin, we are looking at the point from a slightly different angle. It doesn’t mean that we are especially in disagreement. The first thing that I see is that first, we never observe the supply chain directly, so when we say, for example, demand, there is the intermediation of the enterprise systems, which can be very complex. A typical ERP that has three decades is three decades old, and we are going to talk about 2000 tables. Each table has something like 50 to 200 fields, and then if you’re talking to a multinational, they might end up with a semi-nightmarish situation where there is like one different ERP per country, so 40 countries. So first, the input signal is incredibly complex in the sense of just pure IT complexity, not to be taken lightly. So there is first this, I would say, this barrier that even if all they are there, even if the data is correct, the data is not garbage, this is very clean transactional data. The problem is just that it’s exceedingly complex due to the fact that all the systems have never been put in place to measure the demand; they have been put in place to operate the supply chains. So first, we have this, I would say, brutal opacity of the applicative landscape. The second thing that we have is that what we call demand, when you start looking at actual industries, you will see that it’s much more granular. For example, let’s say you have a company that sells electrical materials to B2B clients. The reality is that the order that they get is, they have a client that wants to build a building, and so they are going to pass a big order with potentially thousands of product references, and they will actually schedule the delivery. So, they will say, “We want nine months from now to have all of that being delivered, but we want for the first 500 references to be delivered three months from now, then the first 500, four months from now, etc.”
Nicole Zint: So, Joannes, could you talk to us about how you approach the issue of demand forecasting, and what are the challenges that come with it?
Joannes Vermorel: Yeah, so the granularity of demand can be quite complicated. For instance, if you have a big batch order that requires scheduled deliveries over a period of six months, how do you count the demand? Do you count it when it originates or when it’s going to be delivered? So, obviously, there are plenty of complexities. What I’m saying is that when we say “demand,” it’s not something one-dimensional with a time-series perspective where you can project. Usually, it’s a very multidimensional problem, which can be compounded by the fact that if you introduce your own new type of product that replaces the old generation of products, you’re going to have very aggressive cannibalization effects, just because it’s literally your own superior tech that cannibalizes your own previous tech of products. So, the products are typically very close, and your next generation is an all-rounder better replacement of everything that was before.
David Simchi-Levi: Yes, and that’s why it’s crucial to have the right tools and a well-trained supply chain scientist to tackle this challenge. SQL is one programming language used to extract data, but we need better tools than that to make sense of the vast amount of data we have to handle.
Joannes Vermorel: Exactly. At Lokad, our focus has been on thinking about what sort of tools our supply chain scientists need. We don’t have any AI that can just take data and give us demand forecasts. We need human intelligence to make sense of the data. One question we ask is: What sort of tools do they have? Do they have SQL or better?
Nicole Zint: And what about the plan? How do you approach that?
Joannes Vermorel: Well, Lokad is a very operational company, and we focus on the day-to-day execution of supply chains. What we do nowadays is to make the plan disappear entirely. There are no more plans, no more forecasts. Or, at least, those things still exist as numerical artifacts, but they are completely transient and buried into the system’s data pipeline. The only visible effects are the endgame decisions. For example, what do you buy? What do you produce? Where do you move the stock? Do you move the price up or down? So, if there’s a disruption coming from China, that’s an input, but the fact that it modifies the plan is inconsequential. The only thing people will see are decisions that are steered into something slightly different. And if they look at the driving forces in dollars or euros, they will see that the risks expressed in dollars for certain classes of risk have skyrocketed due to this new information.
Nicole Zint: So, Joannes, what do you think about the impact of disruptions on the supply chain?
Joannes Vermorel: Well, disruptions in the supply chain can cause serious risks to offshore suppliers. It can explode their associated risk and steer all decisions away from them. However, for most companies, the forecast and the plan become a numerical artifact that is quite inconsequential. I mean, there are plenty of other numerical artifacts that are not first-class citizens that don’t capture the interest of the company. I probably need to learn more about what you are doing to give a more substantial comment on what you described.
Prof. David Simchi-Levi: Companies that I have collaborated with face challenges in different parts of the supply chain that may imply different thinking about planning and forecasting. Let me give a very quick example: if you think about fashion retailing, part of the portfolio is a portfolio that I cannot predict in any event. We can generate a forecast, but it’s so unreliable that the supply chain is focusing only on speed. But there are other parts of their portfolio where I can predict very well, and this prediction is used to manage the supply chain entirely. It may be that many of the supply chain executives do not see the prediction, but it is used to motivate where I position inventory, how much inventory position, how do I respond to order. But in the first part, not only prediction is hidden, there is no prediction because it’s so highly unreliable, and the supply chain is mostly focusing on speed. In that part, I think your point is a little bit different, a little bit deeper that even if you have prediction and you have a plan, what you want to demonstrate to the decision making is just the outcome of a specific event, rather than the what contributes to the outcome which is the prediction or the plan. Now, one important element is whether a human decision maker will feel comfortable just looking at the outcome and not understanding what drives the outcome, whether it’s the plan itself or whether it’s the prediction. But I certainly need a little bit more insight about what you guys are doing in order to identify the potential opportunity in this area.
Nicole Zint: David, what emerging algorithms do you see at the moment in supply chain learning and optimization, but also what negative trends do you see equally becoming popular?
Prof. David Simchi-Levi: I think, let me start with the second part, which is the trend around the negative trends. The negative trend around the impact of disruption and volatility that we see in the market is going to be with us for many years to come. This is a new normal, and as a result, companies need to rethink the way…
Nicole Zint: Joannes, you were talking about some negative trends that are impacting clients. Could you expand on that?
Joannes Vermorel: Yes, certainly. From my perspective, there are two problems that are impacting clients. The first problem is that modern enterprise software is incredibly multi-layered. There are layers upon layers of layers, and data is flowing from layer to layer. For modern systems, we’re easily talking about 100-plus layers where the data is flowing. Data science is only adding about 20-plus layers. Just to give you an idea, when you say you want to do data science in Python, the reality is that you don’t do everything in Python. You have layers done in Pandas, layers done in NumPy, layers done inside kit, and so on. A lot of companies are struggling enormously with the fact that over the last couple of decades, the systems have become so multi-layered that every layer is an opportunity to have bugs, regressions, and all sorts of mishaps. This is hindering all the supply chain initiatives in a very brutal and simple way. They try to do something, and at the end of the day, they can’t even get the stock level right just because the AI is flowing through 50-plus systems, and it’s very complicated.
Prof. David Simchi-Levi: Can I add something to that? The implication of what Joannes is saying is that the quality of what people are doing is impacted.
Joannes Vermorel: Yes, that’s correct. The second problem that I see is that some machine learning techniques, such as deep learning, are incredibly technical and add their own piles of layers on top. That becomes very difficult to execute. Certainly, very large companies manage to do that, but it’s very, very difficult. So, I see new classes of algorithmic paradigms that allow us to remove entire classes of layers where we can merge, for example, learning and optimization and the database layer in one. You just remove entire classes of layers so that whatever you want to do in terms of supply chain, you have the actual possibility to do it at scale with IT systems without introducing too much chaos. The reality is that if I bounce back to why companies need so many years to do their things, very frequently, it’s not the pure fancy machine learning bit that requires so much time or the very smart algorithmic part of the system that takes so much time and effort. It’s everything that comes before and everything that comes after, which are very loosely integrated, and you end up with…
Nicole Zint: Sorry to interrupt, but can you clarify what you mean by “everything that comes before and after”?
Joannes Vermorel: Yes, sure. Before the fancy machine learning bit, you need to have proper data pipelines, proper databases, and proper data cleansing. After the machine learning bit, you need to have proper ways to integrate the output of that machine learning bit into the ERP system or the order management system or the WMS system. All these pieces have to be well integrated, and that’s where the challenge is.
Nicole Zint: So Joannes, can you tell us about the complexity of supply chain optimization in terms of data?
Joannes Vermorel: It’s tremendously complicated, I would say, data pipelines. So, the complexity of the logistics of the data actually dwarfs the complexity of the logistics of the physical goods. This is like the vintage viewpoint of a software vendor. But right now, my own observation is that people who are doubling down on things that are just going to let the idea of those companies, the IT complexity skyrocket. That can be a fear response to those dramatic events that took place during the last two years. But that’s not going to make your supply chain more resilient if, in the end, you introduce another class of risk through super complicated blown-ups. Nowadays, I see more and more companies that grind to a halt due to an IT problem that can be a ransomware, but sometimes it’s just like internal bugs.
Nicole Zint: And Professor Simchi-Levi, how do you think technology can be utilized to improve supply chain optimization?
Prof. David Simchi-Levi: If we spoke before the pandemic about the opportunities in using technology, machine learning, optimization for improving business performance and supply chain, people would agree. But executives would be very reluctant to invest in supply chain digitization, digitizing the inter supply chain. Not because they cannot see the benefit, they understand the benefit, but they are worried about the enormous financial investment and the long time it takes to get to the benefit that they are trying to achieve. What the pandemic showed us is that the future is here, that today with the data that is available, we can be more agile and resilient in our supply chains.
Joannes Vermorel: If we want to do more, especially to be smart in terms with complex concepts like resiliency, risk management that do not lend themselves to direct measurements, we need to have tools that are capable of doing that. But we should not develop things that can actually be deployed and put in production in a relatively short amount of time, and that’s really a challenge. The question is that right now, the universal tool that is used to bring a decision to production in 48 hours is Microsoft Excel. And if we literally have something that has all the property that people seek in Excel, which is a decision-making tool that you can use to drive a multi-billion supply chain and take the decision that needs to be taken now, with superior correctness by design, that would be one way of looking at it. That’s definitely the sort of research orientation that we pursue.
Nicole Zint: Professor Simchi-Levi, can you comment on what Joannes just said?
Prof. David Simchi-Levi: Let me build on what Joannes said and connect it to today’s supply chain challenges and IT challenges.
Nicole Zint: So, Joannes, in your opinion, how important is supply chain digitization, and how can it benefit businesses?
Joannes Vermorel: With the technology available to us, we can dramatically improve business performance. We may not be able to achieve all the benefits of full-fledged supply chain digitization, but with a moderate financial investment and a reasonably short period of time, companies can make a huge impact on the bottom line. That’s why, in my opinion, with all the challenges that we have seen, there is one important positive trend. We realize that the future or reality is here, and companies who are thinking about how to move forward need to take this opportunity to start changing and digitizing part of their business to be able to address not yesterday’s challenges but tomorrow’s challenges and opportunities.
Nicole Zint: Professor, do you agree with Joannes’ perspective on the importance of supply chain digitization?
Prof. David Simchi-Levi: Absolutely, I completely agree with Joannes. The benefits of supply chain digitization are significant, and companies that do not embrace it will be left behind. We are seeing companies across all industries adopting digitization and reaping the benefits. From optimizing inventory levels to reducing lead times, supply chain digitization has the potential to transform businesses.
Nicole Zint: And can you give us an example of a company that has successfully implemented supply chain digitization?
Prof. David Simchi-Levi: Sure, a great example is Walmart. Walmart is one of the world’s largest retailers and has been able to leverage supply chain digitization to reduce its operating costs and improve its bottom line. By using data analytics and machine learning algorithms, Walmart has been able to optimize its inventory levels, reduce waste, and improve delivery times.
Nicole Zint: Thank you, Professor, for that example. And thank you both for joining me today for this very interesting discussion on supply chain digitization.