00:00:00 Einleitung: Warum lead times prognostizieren?
00:01:00 Das Problem, lead times als feste Parameter zu behandeln
00:02:21 Warum lead times staerker schwanken, als viele denken
00:04:04 Bimodale Verteilungen: nominelle lead times und explodierende Verzoegerungen
00:05:30 Warum Vertraege die operative Realitaet selten abbilden
00:06:25 Auf dem Weg zu probabilistischen lead-time-Prognosen
00:07:52 Warum Nachfrage die Prognosedebatten dominiert
00:09:33 Das Erbe der Zeitreihen in der Supply-Chain-Literatur
00:12:33 Die Folgen statischer lead times in der Planung
00:13:40 Saisonalitaet von lead times und das Beispiel Chinesisches Neujahr
00:15:33 Wie lead times andere Denkabkuerzungen verstaerken
00:18:16 Warum lead-time-Variabilitaet so folgenreich ist
00:20:20 Fruehe Gewinne durch ein einfaches lead-time-Prognosemodell
00:23:18 Wie man Nachfrage und lead times in einer Entscheidung kombiniert
00:24:00 Der operative Aufwand zur Modellierung von lead times
00:27:52 Erforderliche Daten: Bestellungen, Wareneingaenge und Lieferantenkalender
00:29:34 Lead times zerlegen: Zoll, Transport, Produktion und Haefen
00:33:07 Lieferantenbeziehungen und historische Leistung
00:36:40 Lange lead times versus lead-time-Variabilitaet
00:38:40 Lebensmittelverschwendung und versteckte Kosten schlechter Kompromisse
00:40:31 Warum historische Durchschnitte irrefuehrend sein koennen
00:43:05 Durchschnitte, Sicherheitsbestand und finanzielle Differenz
00:45:33 Das Problem manueller Eingriffe und Zahlenkosmetik
00:48:35 Warum Menschen Prognosen manuell anpassen wollen
00:50:20 Datenqualitaet und Robustheit von lead-time-Historien
00:53:00 Sonderfaelle: E-Commerce und Filialbestand
00:55:43 Wie man Unternehmen von lead-time-Prognosen ueberzeugt
00:57:30 Absurde Annahmen in klassischen Supply-Chain-Modellen
00:59:24 Lead-time-Prognose: eine einfache und guenstige Korrektur
01:00:06 Die versteckten Kosten von lead times und kuenftige LokadTV-Themen
01:02:13 Fazit
Zusammenfassung
Unternehmen tun routinemaessig so, als seien lead times fest, obwohl sie es offensichtlich nicht sind. Dieser Fehler verzerrt von Anfang an Bestands-, Service-, Einkaufs- und Lieferantenentscheidungen. Statt lead times probabilistisch zu prognostizieren, verlassen sich Firmen auf statische Parameter, Puffer und manuelle Korrekturen, die oft Kosten erhoehen, ohne die Rendite zu verbessern. Nach der Nachfrage gehoert der lead time zu den folgenreichsten Unsicherheiten in Supply Chains, erhaelt aber nur einen Bruchteil der Aufmerksamkeit. Das Problem ist weniger technische Schwierigkeit als intellektuelle Traegheit: Manager bevorzugen beruhigende Vereinfachungen, selbst wenn die Erfahrung zeigt, dass sie falsch sind.
Vollstaendiges Transkript
Conor Doherty: So today, Joannes, we’re talking about why people should forecast lead times. Not just the importance of lead times, because I think everyone in the entire supply chain space knows that lead times are important, but how they interact with that fact varies wildly.
I’ve heard you, or sorry, it wasn’t necessarily you who said this, but I’ve seen it circulate within our sort of corner of LinkedIn, that demand forecasting is what the market might want, then lead time forecasting is when you can expect that you’ll be able to service or meet that demand. You’ve also argued, not only in your actual dedicated lead time forecasting lecture, which is actually still quite popular, you argued that basically lead time forecasting is probably the most important source of uncertainty that people are not thinking enough about.
So at a high level, why exactly is lead time forecasting so important? And sub-question, why is it not being talked about and taken more seriously? Joannes Vermorel: The first issue is that lead times usually are approached from a deterministic perspective. So lead time is supposed to be exactly two weeks, or exactly 48 hours, or exactly this or that. And that’s the default stance that you will find in almost all the supply chain books.
They just assume it’s a given parameter. After all, you are in control of your suppliers, you are in control of that, so you should be able to impose a strict delay, and this is it. So that’s the stance that is, literally, I think off the top of my head, in nearly every single academic supply chain book. Lead times are just taken as an input, a given, but also in the field that’s pretty common.
Now, the software, because the thing is that the quasi-totality of the supply chain software that exists on the market, and all the leaders, they’re essentially textbook implementations. They don’t reinvent the wheel. They just follow exactly textbooks. And so if the textbook is saying lead time is a given, it’s a parameter, then the software just follows this principle: the lead time is a parameter, that’s it.
Now, the reason why I object to that is, and that’s what we realized at Lokad back in 2008, I did the exact same thing. Lead time was a parameter, just input, that’s it. I came, a few years later, to the realization that it was empirically unwise, because lead times are varying a lot, and practitioners were doing something that was very counterproductive, which is, “Ah yes, the lead time varies, so we are just going to take a value that is conservative.”
But the problem is that, okay, you are buffering against risk when you do that. You see, if you say, “My lead time varies a little, so I’m going to make it a little bit longer,” but you are still in this deterministic worldview. So what you’re doing is that if you take a conservative lead time estimate, you are making an economic trade-off, but you don’t know what you’re paying for.
Essentially, you say, “Okay, I am willing to say I need more stock,” because if you make your lead time parameter longer, you will need more coverage, and that would mean that you would need to essentially have more inventory, in one form or another. But the problem is that you really don’t know what you’re paying for.
You increase your inventory, and it’s very unclear. And especially when I started to analyze a little bit what lead times look like in practice, lead times tend to be bimodal distributions. So there is one duration that is like the nominal lead time. Let’s say your supplier delivers in three business days, and that’s 95% of the situations. And then the 5% that remains is the supplier is out of stock, or they have a problem, or the transporter, or something, and then the lead time actually explodes.
It’s not three days, it can be three weeks.
Conor Doherty: Okay. Now, which has huge knock-on effects to how you handle demand.
Joannes Vermorel: Yes. And so it doesn’t necessarily make much sense to just increase the coverage. Just imagine you have, you’re ordering strawberries, like an extreme perishable.
Conor Doherty: Yeah, we’re going to get to that later, actually.
Joannes Vermorel: Yeah. It doesn’t make sense to say, “Oh, normally they deliver overnight when I order, but there is a risk of slippage for lead time, so I’m just going to take a five-day coverage.” No, it really does not make sense. It does not.
So what I said is that, okay, it’s not because the supply chain literature doesn’t address the problem that it must not be addressed. And it must be addressed. The reality is that no matter what you say in your contracts, the lead times rarely reflect what is agreed upon. That’s just the reality. For many reasons, the contract might say the lead time, the supplier would say the lead time will be two weeks, but in practice they are much, they’re like three days.
So why would you take a two-week coverage if, in practice, almost always your supplier delivers the things that you just ordered much faster? It doesn’t make sense. So what we observed is that the sort of declarative approach where you say the lead time is this was just factually super wrong, super wrong.
So the solution is, okay, that’s something that is a source of uncertainty and you need to forecast. And when you forecast, just like for the rest of the other forecasts, you need to embrace this probabilistic perspective. And so you end up with a probabilistic lead time forecast. And obviously, it will be very consequential on the decisions that you take.
And I think the fact that lead times are consequential is fairly well understood. The problem is, conceptually, the attitude toward lead time is, again, to just treat them as static, as a known parameter, and thus ignore. Or sometimes, again, in the literature and in some software, you also have a little bit of varying lead time, but it’s done in ways that make zero sense.
For example, applying normal distributions. You will find this in many textbooks, many, many software as well. And it’s a litmus test for complete incompetence on the subject, because if you adopt a normally distributed lead time, it means that you accept positive probabilities for negative lead times. That’s extremely strange, and in fact extremely incorrect. Negative lead times mean you order today and you receive yesterday. It doesn’t make sense. But anyway, a lot of software have implemented this very interesting picture.
Conor Doherty: Well, again, this is what I’m about to respond with. It is somewhat anecdotal, but again, I imagine it aligns broadly with reality, in the sense that demand forecasting is pretty much all you will see people talk about in the supply chain space. That is, if the idea of forecasting is a topic, and we just divide it like a Camembert, like a pie chart, it would be 99% of the discussion is demand, demand, very disagreement about demand, but maybe the slightest sliver is, by the way, maybe lead time forecasting.
So you have this disconnect, is what I’m getting at. You just described the disconnect, and I just want to know, why do you explain the disconnect when people fundamentally, on a day-to-day basis, they know that they vary? They know it’s damaging. We’ll get into the waste later, but there doesn’t even seem to be that much of a drive on the software side to provide the tools that people need.
Joannes Vermorel: So the split is super sharp between demand forecasting and lead time forecasting. I think when I was reviewing on Google Scholar the ratio in the literature, off the top of my head, it was something like, there are over 1 million papers in total for demand forecasting in all variants, and we are talking of something like a thousand for lead times. So the ratio is like 99.9%.
Conor Doherty: Yeah, that’s what I’m saying.
Joannes Vermorel: The imbalance is absolutely massive.
Conor Doherty: So why is that? And you are speculating, but I’m just curious, because it just seems, the disconnect there seems seismic, is what I’m saying.
Joannes Vermorel: Yeah. I think it’s, again, this sort of thing is, back in the ’50s, supply chain textbooks started to discuss time series. And the idea of demand forecasting, time series of demand, was also one of the reasons why it captured so much the interest of everybody, is that on paper, not in reality, on paper, demand can be represented as time series. And forecasting time series was extremely interesting, because if you could do that, you could forecast the stock exchange, the market.
It took a lot of time for the community to understand that the time series of supply chain have nothing to do with the financial time series for the price of commodities, or the price of shares, or whatever. The only thing that they have in common is that both can be visualized as time series. That’s it. That’s the only thing that they have in common.
But nevertheless, due to this attractiveness of, “I can forecast time series,” the reality was, when you think about why we have such an imbalance, it’s because there are tons of people who are focusing on forecasting time series because, again, it got so much interest in the literature, in the supply chain courses, in everything, and even in supply chain software. Again, the quasi-totality of supply chain software will only focus on time series forecasting. That’s all the big ones. They only do that.
So you end up with this self-reinforcing focus that just completely ignores lead times. And because lead time forecasting was also ignored in the books, there were, I guess, very few people that were even thinking in academia it was a challenge. Thus, almost nothing gets published, et cetera, et cetera.
So it is strange, but you can have those semi-accidental focuses where something can be very important and you just have very few people who are actually trying to do anything about it.
Conor Doherty: Well then again, the title of the discussion is supposed to be why people should forecast lead times, and the reasons why this is an incredibly consequential idea. So, to start making it more concrete, we’ve established that people are aware of the importance. We’ve established, I think reasonably, that at least anecdotally the literature has not covered it sufficiently well. And most practitioners we speak to, and again it’s anecdotal, but most that we speak to will tell you something like, “We treat them as, we’re aware it’s not a great idea, but we treat it as static lead times because we don’t know what else to do.”
Okay. In general, talk about the failure mode there. What happens as a consequence in the planning, in the supply chain, when you treat lead times as a static property?
Joannes Vermorel: Then, again, all your trade-offs will be completely off. That’s it. That means that you think that you are well covered, you’re not. You think that you’re tight and maybe reaching stock-out, you’re not. It is a factor that is very consequential for the quality of service.
And the thing that is a little bit extraordinary is that because it is ignored, people are speculating on, again, and focusing on the time series, and that just do not reflect what they will effectively get. And that means that companies, for example, end up with every year the same problems, because let’s say, for example, lead times are seasonal too.
And if you don’t have a tool to actually predict this seasonality, very basic seasonality of lead times, that means that every year you’re going to be surprised by the exact same suppliers who see their lead times being lengthened at the same point of time. For example, Chinese New Year.
Conor Doherty: Yeah.
Joannes Vermorel: And everybody’s surprised. And obviously some practitioners know it. So what they will do is that they will manually tweak the numbers. But doing manual overrides all over the place, SKU by SKU, so that you account for things that can be modeled and solved once and for all, is just madness.
So yes, if I say, do companies who are, for example, ordering from China, are they going to ignore the impact on lead times of Chinese New Year? Probably not. But the person who is doing that is going to do it in a very haphazard manner, by essentially tweaking the purchase orders. It’s very time-consuming. It’s very ad hoc-ish. And again, if this person retires, this knowledge is lost. People need to rediscover that again.
When I say you should be forecasting lead time, it is that this thing should be robotized. That’s it. Lead times, you have the historical data of your lead times, you project them, you end up with those probabilities, and then you combine the probabilities for these probabilistic forecasts of the demand, probabilistic forecasts of the lead times, you combine the two, and then you do the economic scoring for your decisions.
Conor Doherty: So feel free to take a sip of water, because I’m going to need a couple of minutes to get this thought out, but I do want to get your response to it, because I was in conversation with an anonymous practitioner recently about this. And I’m going to paraphrase his thoughts and then get your response, which is that varying lead times, you can’t say whether or not any individual source of uncertainty is the one that’s the killer. But what varying lead times will do is they will exacerbate the potential damage of all the other sort of mental shortcuts that you might take.
So if we say static lead times are a shortcut, it’s just a way of dealing with uncertainty. Whether it’s good or bad, how bad, we’ll just say it’s a shortcut. The thing is, it’s not like it’s the only shortcut that you typically find. In the analogy that he used, it was like hunting for mushrooms after the rain. It’s like once you find one, you’ll typically find other mushrooms. There’s not just one mushroom by itself.
So it’s not like, in this case, supply chain, everything will be state-of-the-art except for that variable. It’s like where you find a static lead times perspective, you’ll also find ABC analysis. You’ll also find service level thinking. You’ll also find FIFO for allocation once it arrives. And the thing is, when lead times vary, those choices, those mental shortcuts that are sort of stacking up, they will become much, much more consequential financially, particularly if you’re in something like fresh food or perishable food, excuse me. Well, all foods are perishable, but you’re dealing with perishable products.
So, for example, you start with an ABC analysis. I don’t know, dairy. It’s like, I need lots of cheese or yogurt, whatever. I’m going to set a service level on that. Again, mental shortcut to make that a bit easier. That’ll be a class A, 98. Again, mental shortcut. How much should I order? Well, I need enough to maintain that service level. Again, mental shortcut. Then lead times, they vary. When the stuff arrives, it’s perishable. You thought you’d have four weeks of shelf life. You actually only have two.
Where does it now go? Well, what was the first store that said it needed a class A SKU? Well, it was the one that’s actually farthest away. First in, first out, off it goes. Shaves off another week, two weeks. So at the start of this long chain of mental shortcuts, you thought you’d have four weeks of shelf life for this product, and then at the end of it, you actually have one. So now you have to immediately do a markdown. And again, another mental shortcut is applied.
So basically what I’m saying is, and to get your response, it’s not necessarily that any one of these mental shortcuts kills you or floods the boat, that they all tend to cluster or covary, and lead times in particular will exacerbate the potential devastation of that. Your thoughts?
Joannes Vermorel: Yeah, absolutely. I mean, first, varying lead times will literally change the magnitude of what you produce or order at any given time by a factor two. It’s very consequential. So again, there are plenty of things that are consequential, but again, lead times, and I would say I would still rank, obviously, future demand as number one, but lead time is a close number two.
So it is very consequential, and there is diminishing returns in the efforts that you make on the modeling side. So here, when people take lead time as constant, it’s super crude, which means that anything that tries to be at least a little bit better helps. Think of it: the constant lead time is as if, on demand, you would pick one number, constant demand, and you would say, “This is it, and we will revise this number in one year.”
Now I give you a slightly better model for demand: moving average of the demand. I just average the demand that I had over the last three months. Way better than just constant demand kept constant for the entire year. It’s still a very simplistic model. It’s still going to be very crappy on seasonality, but at least it will not be wildly off, such as I forecast demand and the product collapses, there is zero demand for six months, and I still project some demand. At least my moving average will be able to capture this sort of thing.
And now if I say I want to capture seasonality with, let’s say, a week-of-the-year effect, I will make my forecast better, but again, diminishing returns. Just having the level captured with an average was giving me, boom, a big improvement. Seasonality was going to give me still quite a big improvement, but smaller, et cetera.
And what I’m saying is, just think of doing the same for lead times. Lead times are very consequential, and if you take them as constant, just think of it as if you were taking the demand as constant. It’s going to be incredibly crap. So maybe you should just have a little bit of following the lead times, at least the recent ones, so that your estimates are not completely off if your supplier over the last six months has been constantly demonstrating longer lead times.
It’s pointless to stay with the wishful thinking of the state that you had before, and then refine with seasonality, refine, et cetera. So what I’m saying is that because you have this diminishing returns, the counterpart is that the initial effort at doing this proper modeling has usually a huge payback.
And here, for lead times, I really see that as a blind spot, a spot where most, let’s say, the quasi-totality of companies that are above half a billion have already spent man-years on demand forecasting. Obviously, when I say man-years, just consider how many hours have been spent over the last two decades on that. And usually for demand forecasting we are talking of man-years, and for lead time usually nothing, minutes of attention, usually complaining about it.
So that’s why I say if you start spending at least, instead of spending nothing, you start to spend a few days to have at least a model that is not completely dysfunctional, you will have enormous gains. And it will also surface the fact that now you have a second source of uncertainty, and you need to have the tooling to combine two sources of uncertainty: demand and lead times.
And if your software and your tools that you have don’t let you do that, you have a big problem. And I think that’s also one of the reasons why people were avoiding that, is that most of the classic literature, uncertainty does not exist. It is treated really as a second-class citizen. And as a result, any kind of tooling that would let you combine various sources of uncertainty is not just weak, the whole thing is entirely absent. It’s not even discussed. There are no tools, no instruments that are proposed for that. And again, things done in the literature cascade in most enterprise software, and in the quasi-totality of supply chain enterprise software, this thing will be completely absent.
Conor Doherty: You brought up exactly what I was about to ask next, which was in my notes: lead time forecasting would not be done in isolation. Arrow, it’s another input to a decision. Now, because I don’t want to assume that people understand what that looks like, because in our heads, we know what that schema looks like, to other people that might sound like an entire other workflow. So as you said, most companies will have teams that have spent work years, or years’ worth of man-hours, on the demand forecasting side.
In order to institute the kind of forecasting you’re talking about for lead times, is that, “I need to hire another team”? What does it look like operationally if I want to have both of those inputs and then merge them?
Joannes Vermorel: Operationally, when we do a Lokad setup for, let’s say, a $1 billion company, just take a baseline, a state-of-the-art forecasting model is just going to be something like a few hundred lines of code. That’s it, for demand. And maybe sometimes it’s just 100 lines of code. And lead times, even less. And I’m serious when I say state-of-the-art.
When we look at the competition, the M5, that was a Walmart data set. Some people at Lokad landed number one worldwide at the SKU level with a model that can be written with literally 50 lines of code. And yes, there were plenty of people who took part in this competition that had models that were hyper-complex, tens of thousands of lines of code, models with millions of parameters, and yet at the SKU level we were number one with just a relatively simple parametric approach.
So bottom line is, usually those man-years are completely wasted. There is no, those companies are way past the diminishing returns. It’s just like there are no returns. It’s a complete waste of time. It’s a bureaucracy that produces nothing but PowerPoints. The search for a greater accuracy stalled decades ago.
Usually, very frequently, when we start working with one of those typical companies, the reason why I say $1 billion, it’s typically a company that is at least three or four decades old. So those things have been done for a long time, and their capacity to predict the future has been super stagnant for decades. And those extra efforts, they are just noise. They are not making progress.
So typically the way we approach that is to say we need to do the things right and have something state-of-the-art at the demand level, which is typically just a few days of work, not really much. And then you need to switch to the other sources of uncertainty, and that’s going to be lead times. That’s going to be the prices of your supplier.
Sometimes, for example, the lead times are a little bit tricky because depending on the quantity you order, you don’t have the same lead time. So there is this sort of self-prophetic effect where if you pass a massive order, then your supplier struggles and it takes longer.
Conor Doherty: So you’re not doing any per-SKU ordering.
Joannes Vermorel: Yeah, I mean, it really depends if you’re passing an order and it represents 0.01% of the volume of your supplier, or if you pass an order and you represent like 60% of the instant quarter volume for the supplier. Very different situations.
And you have to, again, factor those elements that can be also forecast. If you’re e-commerce, the returns. So you have all those. Lead times are just one of the uncertainties. And in the end, my take is that if you have the correct tools and proper know-how, the amount of time that people should spend, and again I’m talking like a $1 billion company, per year on probabilistic modeling is a matter of weeks. It’s just a few weeks per year.
Conor Doherty: Okay, I’d actually written down, because I wanted to ask you, because I want to perceive this from a potential end user’s perspective, or somebody who wants to adopt this kind of perspective. They’re thinking, “Oh, this sounds great, but what kind of data is required of me?” So I have just a list of things. You can just tell me yes or no, or how critical you think these data sources would be to institute the kind of change that we’re talking about.
Obviously, historical purchase orders.
Joannes Vermorel: Yes, obviously.
Conor Doherty: Receipt dates.
Joannes Vermorel: Usually yes. You need to know when you ordered and when the stuff actually landed in your warehouse, or in your store, or wherever you want the thing to land.
Conor Doherty: Supplier calendars, nice to have?
Joannes Vermorel: Nice to have, but depends. Again, if you’re FMCG, usually suppliers have very simple rules. You order on Mondays, they give you a day, and you stick with that. Again, it varies very much, but there are few industries where it’s very complicated and you need to have exactly the calendar of your suppliers.
For many verticals, it is not that important. It will be, for example, more important to have granularity in what was the shipment schedule that was demanded initially. For example, if you are, let’s say, fashion and textile, you say, “I order 2,000 units, but I want you to send me 500 and then 500 and 500.” That is information that would be more important.
Conor Doherty: This is an interesting one, because it opens up another dimension to the idea of probabilistic forecasting: customs delays, and especially relevant today.
Joannes Vermorel: Yes.
Conor Doherty: Okay, so then how does that, is that a separate input to the lead time forecast? Because it’s like a Matryoshka doll effect, where I take a thing out and there’s another one inside, and then there’s another one inside. So it just seems like this is never-ending, really.
Joannes Vermorel: Yeah. I mean, any duration can be decomposed in sub-durations. So if you want to have predictive probabilistic modeling of your lead times, I was saying you should grow into progressive sophistication to capture more, to make it more accurate.
First, you’re going to just have the baseline. Then you will start to include the cyclicities. And then, if you can collect the data, you will start to try to decompose your lead times into their components, and then you will try to forecast the components independently.
For example, if you order from China or from Vietnam, your very long 10-week lead time will be decomposed into how much time your supplier has to wait to have their own inventory ready at hand to start the production, then the duration of production, then the duration of transport toward a port, assuming it’s a port, then the duration to wait for a ship, because depending on how much you price your stuff, it might be sitting for a while, and then the shipping, the fact that the cargo ships move, which depends on the weather, and then on the final ports you have to see how much time it will wait, et cetera, et cetera.
So obviously you have all those steps. If you can measure them independently, the interesting thing is that the sort of probabilistic modeling that you will have, you will have piecewise elements that will be much more accurate, because you will be able to have a modeling where each stage is modeled more accurately.
But if you can’t, that’s fine. It is a refinement. This is not strictly necessary. But at some point, it will be a ballpark assessment, which is, are the probabilistic views that we have of our lead times too coarse? Do I need to delve into the fine print? And again, it will really be a judgment call depending on the situation. Sometimes the stuff that you want to receive from, let’s say, Vietnam is stuck for three months in customs. Then probably you want to factor the customs explicitly.
So it really depends. And then there is the question of whether you have access to the data at all. If you don’t have access to data, it’s completely opaque, then so be it.
Conor Doherty: It occurs to me, one of the examples that we had used, we were talking about the perishable food. If we’re talking about cheese, yogurt, you gave strawberries earlier. If you’re forecasting demand and you’re, let’s say, your supermarket, you don’t really have the same relationship with any individual client. You don’t have a service level agreement with, I’m Monoprix in France, I don’t have a service level agreement with you, Joannes. It’s not that kind of arrangement I have. So, and I can kind of factor that into my decision-making.
That said, when you’re talking about forecasting lead times, there is a relationship with your supplier that you have to consider. And that’s what kind of separates and makes it a little bit more politically difficult. Is there a relationship management dimension to this? Like, you need presumably to be keeping somewhat detailed historical scorecards on your supplier’s performance, for example. Yes. But how do you work with that? Because again, if it turns out that your supplier is kind of actually statistically crappy, now it’s a very different conversation to, “I didn’t have strawberries for Giannis on Monday. Well, there’ll be 10,000 more like him tomorrow.”
Joannes Vermorel: That’s a point. This thinking of “we can control the supplier” is exactly what leads, I think, the literature and then companies into “lead times are just going to be a parameter.”
Conor Doherty: Mhm.
Joannes Vermorel: And if the supplier is not compliant, we are just going to be very, very severe and yell over the phone to the supplier, say, “You have to comply. It was seven days in the contract. Do it.” But again, the reality is that if your supplier is crap, then probably you need to change and find a new supplier.
That will take time. That will maybe take three months, six months, maybe a year. So right now, when you model the lead time, you just have to assume your supplier is just going to be as usual. Even if you want to have the long-term view and bring improvements, change, the reality is that it’s not under your control. And if a supplier has been demonstrating consistently that 5% of the time there were problems, and then they were not compliant with the on-time-and-in-full sort of things that you wanted to have, then you should not, it’s wishful thinking to just assume that it will not happen. It will.
It’s just that there is this risk, and this risk needs to be accounted for. And the fact that you want to improve and reduce this risk by having a tighter process is a separate track. You should not confuse the two. So you need to look at the reality of your suppliers, of your delays, as they are, and then separately say, work on re-engineering the processes and everything to make them tighter. But that’s two different tracks.
For the first option, it’s really what I am ordering now, and I need to take into account the proper factors: uncertain demand, uncertain lead times. And then there is the other discussion, which is how can I have, generally speaking, better suppliers, more reliable suppliers? And also you will start to factor the fact that reliability has a price. Maybe you can get a more reliable supplier, but what if this supplier is a few percent more expensive than the unreliable suppliers? It’s a trade-off.
Conor Doherty: Actually, it occurs to me as a follow-up, from your perspective, and again it’s just a judgment call, but from the decision-making perspective, is lead time variability more concerning to you than long lead times themselves? So, given, if you know it’s going to be a long lead time, 15 weeks, but then you tell me, “Well, you’re ordering something, and the variability of it, we have a bimodal distribution on this decision. Ninety percent confidence it’ll be here in three weeks, or it could take five months because this strait or route closes.” Which is more alarming? Again, that’s just a question.
Joannes Vermorel: Obviously variability, and again variability because its effects are surprising. When you do an economic calculation, you don’t know which way it will go. You might say, “Oh, variable lead times, I need to cover myself more.” Not necessarily. Sometimes you want less.
An example is, imagine you are doing fashion and you want T-shirts for the summer, but your supplier is very unreliable and lead times are very varying. What you don’t want is to receive those summer beach T-shirts in September.
Conor Doherty: Exactly.
Joannes Vermorel: So sometimes the lead time variability is something where you say, “Oh, maybe I will miss the proper period and it’s not worth it.” So that’s why I would say variability is, I think, a little bit more alarming, just because if you don’t take that into account, it can really change your results, and you don’t even know in which direction the economic trade-off will flow. It can be order more or order less. People would immediately assume that more variability means bigger buffer means extra, but not necessarily, not necessarily.
Conor Doherty: Well, it’s interesting because, again, talking about the economic trade-offs, I don’t have figures to hand for fast fashion, like what percentage, what is the cost of just dead stock, for example. But I did do a bit of research before this about food, and I came across two sources. I know one of them, one of them I’m positive about, it’s from the United Nations Food and Agriculture Organization, that about 13% of, sorry, in terms of inventory, food inventory, there’s about a 13% total write-off at the supply chain level. So just to get from point A to point B, you can assume about 13% will just be written off immediately upon arrival as waste. After that, so what’s left, so there’s 87%, a further fifth to a quarter will just be written off through waste due to the mental shortcuts we were talking about.
Anyone who’s questioning those figures, I am absolutely certain on them. Feel free to ChatGPT that right now. But again, what we’re talking about here is not trivial. We have a series, and we’ll come back to this topic, hidden costs, but that is a staggering amount of waste. And again, it is directly attributable to mental shortcuts, at least some of them are, excuse me, some of them are within your control. You can’t stop the strait from being closed. You can’t do that. But you can go, well, how am I classifying my inventory? How much am I going to order? When do I think it’s going to arrive? What am I planning for? What assumptions am I baking into my own decisions that I do control? When it arrives, what do I do with that?
These are within your control, and they contribute to a statistically significant amount of write-off, even expressed against actual turnover.
Joannes Vermorel: Yes, huge amounts. And that’s where, again, the probabilistic approach really shines, because with lead times you end up with surprising results. Let’s assume that you say, “Oh, I’m not going to do those probabilities. It’s just going to be the average lead time.”
Conor Doherty: I actually have a question about that.
Joannes Vermorel: The problem is, if you have sometimes products that never arrive, your average lead time is infinite. That’s literally. And you can say, “Oh, I’m just going to say that when it never arrives, the lead time is one year.” That doesn’t make sense. It never arrives.
So yes, if you say when it never arrives, I average out this thing taking one year, but why not two years or 10 years? So bottom line is you end up with a distribution that has no mean. I know it may sound strange to this audience, but there are plenty of mathematical distributions that don’t have a mean. You don’t have a way to compute the average. The average doesn’t make any mathematical sense.
And that’s the case for many situations for lead times where, when there is a certain probability that products never arrive, and again, that’s, you were talking about situations that are very alarming and whatnot, that means that, for example, you need to take that into account because when you do a calculation for what I should be ordering next, you need to take into account what is already in flight, what is incoming.
And that means that if you have stuff that has a probability, even if it’s low, of never arriving, it can substantially impact what you’re about to order now. And that’s the sort of thing where lead times are not, I would say, extremely complicated, certainly compared to modeling the demand. Typically, the model, when we implement a model in terms of lines of code, I would say rule of thumb, the lead time forecasting probabilistic forecasting model is going to be half of the lines of code of the one for demand.
But nevertheless, it comes with its own twists, and you have this sort of situation where you need to be careful about the fact that, for example, you can have infinite times, which is fine. It just means that you need to have the instruments that don’t go crazy on those situations, and that’s it.
Conor Doherty: Okay. This next question, I want to preface it by saying I know what the reaction will be, but I also want to make it clear this came from someone who is a fan. So they’re asking the question from within their perspective. It is not emblematic of their actual feelings.
So, quote, “We already track supplier lead times. We use the historical average” — you’ve just commented — “We use the historical average, maybe adding a bit of safety stock on top of that. Please ask Joannes to comment on the delta plus financial delta between that approach and what you can expect from what you just described.”
Joannes Vermorel: It’s enormous. Again, that is FMCG as well.
Conor Doherty: Yes.
Joannes Vermorel: I mean, go back to the example I gave where the demand will actually collapse because it’s too late. So you add a buffer for the wrong thing. Imagine you’re in a situation where, in case the delivery is too late, the demand, which is seasonal, is already gone. And now you add a buffer. That was a bad situation and you’re making it worse.
And that’s why I say the problem of just buffering is that this is not an economic perspective. You don’t know what you’re doing. You can always say, I take a purchase order and I just decide to order 20% more. That’s a shortcut that we’re talking about. But then if you want to be rational about it, you need to think of rate of return. Am I doing something that will increase the rate of return for my company or not? That is the essence of the game.
And here, when you just add buffers like that without knowing in which way you’re actually steering the rate of return, the problem is that you don’t know. And that’s the point, is that when you do it, essentially what you’re doing is random guesswork. It’s extremely uneducated as a sort of correction.
And that’s the problem, is that if you pile up tons of largely uneducated number fudging, then why should those things add value? There is a psychological distortion. If I take one number that is coming out of a numerical recipe and then I ask people, “Do you approve?” there would be x% of the people who approve that. And now that’s experiment number one. And now I say you have the opportunity to fudge this number up or down. Okay, people do that. And now you say, “Do you approve?” And the percentage of people who, in the second situation, will approve will be much higher.
There is this sort of, it’s a little bit the IKEA effect. People will feel a lot more comfortable with something where they have assembled themselves, contributed, doing something. But it is irrational. You’re not making a purchase order better just because, on your gut feeling, you push the number a little bit up or down. That is a psychological problem, is that this sentiment of, “Oh, now that I’ve tweaked the number, it’s a little bit mine, I feel more at ease with it,” is nonsense. It’s not rational.
And very frequently it creates all the sort of incorrect busywork in the company where people spend a lot of time doing this micromanagement that, when you look at it from an economic perspective, my experience was it is almost invariably a net destruction of value.
Conor Doherty: So just on that point, just to provide some context that supports what you just said from outside of Lokad, I’m trying to recall the actual name of the paper, but I do know it was in 2023. It was Robert Fildes, Paul Goodwin, and I think it’s Frank D… I’ve never met the man, I’ve never read his name out loud. It was those three authors, investigated, they were talking about forecast value added, and do manual overrides increase, sorry, what are reasons, investigating the reasons, and do manual overrides increase or decrease? Do they make the forecast better? And they found that no. Again, I’m summarizing. I’m sure they might challenge the summary, but overall, no.
But most importantly, and most interestingly, the justifications that people gave were not things that were actually present in the data. They’re not things that you could look at and go, “Oh, well, actually I’m pointing at that number. That’s why it went up and went down.” What actually led to the overrides was, “Well, what did I do last time?” Things that were exogenous, things that were outside of the observable data that should influence.
But it does raise an interesting question, which is most people, in that analogy that you gave, most people would feel somewhat confident or comfortable to start tweaking a demand forecast. Do you see the same kind of issue with lead time forecasting, where people will want to get their fingers in the pie? Because I don’t even know how that would work. Like, “No, no, I think that should be 21 days, not 25 days.”
Joannes Vermorel: No, I mean, people can do that, but generally it’s, no, the amount of, again, on this front there were not 20 years of S&OP process to brainwash people into the fact that they should be tweaking those numbers. Usually, the main challenge is to bring companies to the conclusion that they should abandon those static lead times fixed once for all, as if they were the truth, to go for something that represents the uncertainty.
Usually the reaction is, “No, I mean this variability is a defect. We should just, if you start having a model that shows that the lead time is varying, you should just make a phone call to the supplier and say, ‘Stop it. Just be compliant, on time, in full, and that’s it.’” But again, it’s wishful thinking.
If the supplier has not, for years, given you a perfectly reliable, non-varying lead time, there is no reason that after one more phone call they will start doing that. It can be just because the cost of reliability is just not worth it for them.
Conor Doherty: Well, before I push on, I do want to just say, if I have mispronounced anyone’s surname when I was citing that, do please forgive me. I’ll issue a correction when we post this online if I’m incorrect.
But to push on just a little bit, I know that when we’ve spoken before about master data for optimization projects in general, and most people, when they hear optimization projects, they think demand forecasting. To them it’s a byword. These are synonyms. “I’m optimizing my supply chain. I’m optimizing demand.” And they think of master data in terms of that. Fine, that’s not the topic for today.
But how, so you’ve been very forgiving of, like, look, your data as it lies.
Joannes Vermorel: Yes.
Conor Doherty: Just give it to us, or give it to whoever is sorting it out for you.
Joannes Vermorel: Whoever your vendor is, they should handle that.
Conor Doherty: I presume the exact same holds for lead time forecasting, because, as we pointed out, the quality of data and the robustness of record-keeping for lead times might not actually be anywhere near as robust as the already sort of wishy-washy historical transactional data.
Joannes Vermorel: I mean, it is historical transactional data. If you don’t track what your suppliers delivered, because if a supplier says, “I delivered 1,000 units to your warehouse,” the first thing that you will have to do is recount. So usually the transactions are quite good. Things are sent and then counted, because it’s trust but verify.
Conor Doherty: So supplier performance, I mean, how robust is the average company’s records of that?
Joannes Vermorel: Again, I’m just talking about the what is. You know when you order, you know when you receive stuff, and that’s it. There are a few tricky cases, but it’s more like for e-commerce, when you want to know when your B2C client received the good. Then you have a problem, is that you get this data from the transporter, but you don’t know exactly if it’s correct. The transporter can say, “Ah, delivered.”
Conor Doherty: As happens in France.
Joannes Vermorel: Yes, yes. But in fact it’s not. So okay, e-commerce here, it’s an edge case, and then the lead time is something a little bit different, because it’s not a lead time in the sense of how much time do I need to organize my production or my purchasing. It is more about the promise that I’m going to do to my customer, because you want to make a promise. And here you need to give something that is like an upper envelope so that you don’t disappoint your end customer.
So there are a few edge cases where indeed establishing lead times ends up being quite tricky. This is not the only situation. There are, for example, the stock levels in grocery stores also involve the same sort of problem. If your electronic system says there are three units in stock, it cannot say for sure that one of the three might be damaged or misplaced, and thus the perception of the clients might be it’s not two units, three units that are left on the shelf. It can be zero because one is damaged and the two others are misplaced. And so the actual perception is there are zero units left, while your transactional system says there are three.
But overall, I would say the situation for lead times is not too bad. Almost invariably, companies that have above, let’s say, $50 million of annual turnover, those things are sorted out. The quality of the data with regards to lead times is usually very good. And for companies that are above half a billion, it’s usually excellent.
Conor Doherty: Plus, I’m paraphrasing something you said before, echoing something you said before, which is that a competent vendor is going to take care of that for you. It’s not going to be something that you have to sort out on your end.
Joannes Vermorel: Yes. But also, you see, again, there is a Darwinism at play, which is companies above half a billion a year who were not able to keep track, at least a little bit, of the delays for the deliveries or for the productions went bankrupt.
It’s a little bit like if you can’t keep track of how much you’ve paid your supplier, you will go bankrupt because you will be paying certain suppliers twice, just mistakes. If you can’t track if your customers have paid, then they will have things that will never be paid and you will go bankrupt. So it’s part of the reason why the transactional data tend to be very high quality. It’s just Darwinism. The companies who did not manage to be good at this game of having clean transactional data, they just went bankrupt. It’s like a self-selecting mechanism.
And so companies that reach a certain size, they are survivors. They managed to pass those vessels. So that’s why, usually, again, my experience is that the transactional data for companies above, let’s say, $50 million per year is usually excellent, and above half a billion it’s usually extremely good. Really, it’s beyond excellent, it’s quasi-perfect. Doesn’t mean simple. Doesn’t mean not messy.
Conor Doherty: Well, historically, based on the feedback we’ve often received from people who actually agree with us, convincing people to think that they should be forecasting demand probabilistically is a bit of a mini battle, and that’s putting it nicely. A bit of a mini battle, a war of attrition. Sisyphean is an adjective that I have heard more than once.
But what we’re advocating for today is, okay, so shove that or add to that lead time forecasting and probabilistic forecasting. So what is the, to avoid extinction, what is the pitch to get people to at the very least start thinking that lead times are things that you should be forecasting? Whether or not you get into probabilistic, okay, fine, but it’s a thing that you should be thinking more about.
Joannes Vermorel: Anything where the future is uncertain, you need to forecast all the sources of uncertainties, and that’s it. Obviously, you want to rank that by sort of importance. As you said, demand is number one. Lead time is number two, marginally below number one.
Conor Doherty: Yes.
Joannes Vermorel: But yeah, exactly. Demand is number one, lead time is number two. And then if you’re e-commerce, returns would be number three.
Conor Doherty: So it’s not just a source, it’s a critical one, is what we’re saying.
Joannes Vermorel: Number three would be probably varying buy prices, et cetera. So again, usually those sources of uncertainty, for most companies, it’s going to be a handful, like five or less. It’s not very complicated.
And again, the benefit of at least spending a little bit of time so that you don’t make a crazy assumption such as it’s constant. Just imagine you’re an FMCG company and you say, “Oh yeah, no problem. Price of oil constant. Price of energy constant.” Now you’ve been disproven. Your assumption has been disproven.
The interesting thing is that usually when you point out the craziness of those assumptions to executives, they would say, “I can’t believe we have been doing that.”
Conor Doherty: Yeah.
Joannes Vermorel: And I usually say, “Yeah, I know. It’s barely believable, but you did.” And it’s just, I know you do. Again, for this audience, fun fact, the EOQ formula, the Wilson formula, it assumed constant future demand, and it was invented for essentially shipment by the sea, when people had to arrange a warship.
So that was really the explicit case of, “I will have a warship for me,” and that’s where the reasoning behind this modeling comes from. And we have seen companies in aviation doing that. Nothing fits. They are not going to arrange a whole airplane, and then they can take just a fraction of the cargo space of the airplane. Nothing fits.
And yet, it happened for us more than once that we saw a prospect, turned client later on, where we discovered that for decades they had been using a formula that was absolutely not fitting what they were doing. So for me, that’s a little bit of the insanity. And the upside is that the fix is relatively cheap and short and not really difficult.
All in all, forecasting lead times, that would maybe be a message at the end, is it’s not difficult, even if you want to do it in a probabilistic fashion. I give an example in the lecture. You can do that with a model that is like 10 lines of code, and it’s already decent.
Conor Doherty: Yeah. Fun fact, I believe that was the best-performing lecture that you ever did. I checked, just so you know, I’ve checked that on YouTube. I do believe that is the most popular one. As I said, I will drop that in the comments if we’re watching this on YouTube and LinkedIn, the link to that, 5.3, I believe it is, lead time forecasting, definitely something people should check out.
Also, we will definitely come back to this topic in a live format, and it’ll definitely be the hidden costs of lead times, because again, when I spoke about this idea to other people in our orbit, they were very interested, and a lot of these questions are kind of formulations, and the examples are formulations, because a lot of people, and you said that, a lot of people do know, practitioners know. That’s the thing. It’s probably, more than demand, one of the most frustrating things for people, because they know that it makes no sense, and they’ll say, “Yeah, my lead times are killing me. What do you think of it?” “Well, we treat them as static, obviously, and we know that’s wrong.”
And it’s like they beat me to the sentence. Like, “We know that that’s wrong, by the way, but that’s why we’re having the conversation.” It has this kind of vibe of insanity. “Yeah, it doesn’t make sense, but we do it anyway.”
Joannes Vermorel: Okay, okay. And that’s very strange. That would be, again, I’m struggling to think of any other domain where there is even 1% of this sort of insanity that I see in supply chain, where people say, “Oh yes, we do a super fancy mathematical model, and yeah, it’s completely bogus, but why not?”
And the worst thing is that we are talking of undertakings where there are a lot of resources. Usually it is not like you’re talking of, let’s say, the barber shop, where it’s just one person with just a tiny budget in time, effort, et cetera. No, no, it’s like, “Yeah, over the last decade we invested, in total, $20 million-plus into this setup, and no, we did not manage for this budget to squeeze one week of engineering effort to actually have a decent lead time view.” That feels super insane.
And again, it is strange. It is very strange.
Conor Doherty: Well, it is a topic that we will come back to, as I said, certainly in a live format, and hopefully soon. But, Joannes, we’ve been talking for quite a while. I have no further questions. Thank you very much for your time. And to everyone for watching, thank you very much. We’ll see you soon.
As I always say, if you want to continue talking to Joannes and me, you can reach out to us on LinkedIn, or, say it with me, send us an email at contact@lokad.com. And with that, we’ll see you next week.
Get back to work.
