Simplifying Probabilistic Forecasting

This podcast was designed for video or audio enjoyment. The following transcript was created using a combination of automated and human transcription. Please check the audio version before quoting from the transcript.

Mary Vasile: Hi, everyone, welcome back to the Be Ready for Anything podcast. I’m your host, Mary Vasile, and today I’m delighted to welcome Stefan de Kok who is going to shed some light on probabilistic forecasting. Thank you for joining me, Stefan. Why don’t you tell me a little bit about yourself and your experience in supply chain.

Stefan de Kok: Hi, Mary. Thank you for having me. I’m an applied mathematician by education and almost 25 years of career I’ve spent pretty much all in supply chain – functional consulting, software consulting, software design, software development, as well as a little bit of hand-on. And nowadays I speak and write about the concepts. I’ve worked for quite a few software companies, including ToolsGroup, at some point in the past, and nowadays I have my own software startup, Wahupa, in that area.

Mary Vasile: Great! Well, we’re so excited to have you, and if there’s anyone that can shed some light on probabilistic forecasting, I think I’ve come to the right person. So probabilistic forecasting has become more mainstream in the supply chain space in recent years. But there are still a lot of questions about what it is. So what have you found to be the easiest way to define probabilistic methods?

Stefan de Kok: Yeah, so it’s difficult. It’s, you know, you’re trying to explain something mathematical. It’s never going to be easy. The way I tend to do it is I explain that the reality is uncertain. And when uncertainty exists, probabilistic methods are really the only accurate way to look at that reality. So internally probabilistic methods use all the uncertainty in all their calculations and express all the potential outcomes as probabilities that may occur. If you compare that to a deterministic method, they use individual single point quantities such as average expected sales or given period for a given item in all their internal calculations.

So as an analogy, what I use is imagine you’re going to the casino and you’re going to play a game of craps. And this is a game played with two dice and the total number of the dots on top determine your outcome. The deterministic, old-fashioned way, I would say, of looking at it is you’re going to throw a seven every time. Why? Because that’s the expected average, whereas the probabilistic way would look at all the odds that might occur.

So with the probabilistic perspective, you have a chance of actually winning some money or losing, still, but it depends on your skill and some luck, whereas the traditional way, you really don’t have a choice. You’re going to lose your shirt no matter what. And that, I think, is an analogy that works in business as well. The probabilistic approach gives the decision-maker much richer information upon which to make informed decisions that simply doesn’t exist if you think every outcome is a single number.

Mary Vasile: Now, you mention on your blog that this concept is finally starting to gain some traction. What would you say is causing the increased use and awareness of these methods?

Stefan de Kok: I think there’s a couple of factors. One is a lot more blogging is happening by various parties, ToolsGroup and myself included, but also there’s more alignment on terminology. So in the old days, everyone used different terms. If you went to academia, they call it a density forecast. ToolsGroup was calling it demand modeling. Other companies were calling it quantitative methods. So there really wasn’t a single way of looking at it. And at some point, it all started coming together. I’d like to think I had a little role in that, but basically, all started to realize that they were all just flavors of the same thing and it’s all just probabilistic. And once we got that alignment, it became an easier message and more voices out there doing it.

And of course, businesses are getting more uncertain these days, especially this last year. We’ve definitely seen a speed-up due to the pandemic. And people are looking into things such as risk management, resilience, and naturally, then the probabilistic approaches tend to get more views when that happens.

Mary Vasile: So you mentioned now business is more uncertain and the pandemic has definitely driven that point home, but in the face of uncertainty then, why probabilistic forecasting? What are some of the greatest advantages to these methods?

Stefan de Kok: Well, first of all, what I just mentioned, the richer information to make the information on- to make the decisions, sorry. And there’s also the information that you’re getting is more accurate.  In the traditional way, you might have confidence ranges. But these confidence ranges are just as biased, maybe even more biased, than the point forecasts and plans that they are around. So having more accurate information, having these distributions that are true reflection or close reflection of the real world that allows you to calculate much better buffers and plans using output of probabilistic forecasts and probabilistic plans.

And I think the final one, and that is one where there’s not much happening yet, is the ability to distill signal from noise and natural variability from error. That also requires us to change the way we measure quality to really get the full benefit of that. I don’t see that happening a lot yet, but that’s certainly coming. I’m pushing hard for that.

Mary Vasile: Now, benefits like these are enticing to any planner or company, but can you tell me what sort of, what sort of business environments are best suited to probabilistic forecasting methods?

Stefan de Kok: I would, I would say environments where recurring variability is the dominant factor, right. And recurring is the operative word there. The moment you have patterns in history, you can determine patterns for future out of it. Not necessarily history repeats itself, but things that occurred in the past, if they were to occur again in the future, what’s the pattern in there. Anything with lots of demand to supply history, such as supply chains are a great area for that. Certain activities when they occur like promotions, similarly.

But if you compare that to environments where that is not the dominant factor, such as where maybe human decisions are a factor- the stock market- or you can determine some kind of uncertainty. But if the market just flips and does something on a dime, then whatever ranges of uncertainty you determine prior are out the window. So those are not really the best kind of area for it.

I think it would still yield at least equal or better results than not using probabilistic, but it’s not where it shines.

Mary Vasile: Now, Stefan, you’ve championed probabilistic forecasting for a while, and in your experience, what are some of the most common misconceptions about it?

Stefan de Kok: Yeah, so that’s one- there’s a whole list of it. But I think the number one I encounter these days, as the term probabilistic forecasting has become more popular, more vendors, more consultants are on the bandwagon and they’re claiming that they’re doing it. And so you hear this term “range forecast” a lot. And they say, oh, we’re probabilistic because it’s range forecasts. Well, that is not probabilistic. That is just the old-fashioned confidence ranges that we have on statistical forecast with a new name on it.

So range forecast. I see it as the kiddy bumpers on the bowling lane, right. Sure, you can prevent in almost all cases that the ball goes in the gutter, but it’s still up to you to decide where in the lane and how many pins you’re going to hit. So that’s a range forecast. It’s not really a true probabilistic forecast.

The other one is that A.I. and machine learning are probabilistic and they are not. Right, some of them, maybe five percent, have some kind of probabilistic parts to them, like there’s some stochastic methods, which is a form of probabilistic. Most that I’ve seen still assume a normal distribution. So the probabilistic approach says, no, we’re not going to take the easy distribution. We’re going to find the most accurate distribution. So if you’re not doing that, if you’re already assuming that naive normal distribution, in my book, it’s not probabilistic.

And then the final one, I think, and that’s what’s a real barrier to a lot of companies even considering probabilistic methods, is they think it needs a lot of data and a lot of computing power. And I think both our companies have discovered that that is absolutely not the case. The same data that you would use for a non-probabilistic method is the same data you would use for a probabilistic one; you just look at it differently.

Mary Vasile: And with that, Stefan, we come to my final question for you. What advice do you have for supply chain practitioners for achieving success with probabilistic forecasting?

Stefan de Kok: So my advice to everyone on all topics when it comes to anything in science and life is keep an open mind. And that applies here, too. So, and when it comes to probabilistic forecasting, the tricky part is the math is very sophisticated, right? So unless you are a die-hard mathematician, you’re not really going to get all of it. But as a planner or a decision-maker, you do not need to be that kind of die-hard mathematician, right? You do not need to understand exactly how the math works, but you do need to understand the concept. You need to understand what is uncertainty and why probabilistic perspective is the correct way to look at the uncertainty that you have in real life. And of course, you need to understand your own business. And that’s something that a consultant doesn’t necessarily know or a software vendor like us. We don’t necessarily know your business, but that’s that crucial part you as a subject matter expert would add to it. And then it is up to us, the software vendors, to provide the very rich visual information that bridge the gap between the math and the understanding of the user and use that to build trust.

So I like to compare that with a car. You don’t need to be a mechanic to drive your car and to feel trust that you and your family are safe and your lives are not in tremendous danger as you get from point A to B, right. So over the years, these cars have gotten more sophisticated. And our trust in cars has grown over the years. And I think it’s the same with these probabilistic methods. They’re more sophisticated, but our trust will ultimately be much greater than the statistical ones because they just tend to be right more often.

Mary Vasile: Well, Stefan, thank you so much for joining me today. These are fabulous insights and I really appreciate your taking the time.

Stefan de Kok: You’re very welcome.

Mary Vasile: And thank you, everyone, for joining us. You can find this and all episodes of our podcast at You can also listen on Spotify for the Be Ready for Anything podcast. Thank you so much and see you next time. Bye.