00:00:07 The importance of forecasting trends.
00:01:36 The limitations of using social media data for supply chain forecasting.
00:04:26 The disconnect between social media trends and SKU-level supply chain data.
00:06:37 Challenges in finding early signals in social media data.
00:07:20 Using company data for better forecasting and the importance of stable forecasting models.
00:10:07 The difficulty of predicting extreme events like COVID-19 and the importance of probabilistic forecasting.
00:13:42 The slow and manual process of turning forecasts into decisions in most industries.
00:15:00 The sluggish nature of supply chains and the inefficiency of acting on early signals.
00:15:53 The need for robotization of data pipelines to generate decisions more efficiently.
00:17:07 The limitations of forecasting technology and embracing uncertainty.
00:19:35 The role of supply chain scientists and human insight in decision-making.
00:20:31 Assessing the potential time savings of better forecasting technology.
00:22:00 Prioritizing automation in decision-making over advanced forecasting.
Kieran Chandler and Joannes Vermorel discuss the challenges and potential of using external data sources like social media for predicting supply chain trends. Vermorel argues that while valuable information may exist on social media, it is challenging to match the data with specific products, and the quality and context of the data are often unclear. Instead, he recommends focusing on internal data for sharper future views and adopting a probabilistic approach to prepare for extreme events. Vermorel emphasizes the need for companies to automate their decision-making processes to improve their agility and better respond to demand fluctuations.
The discussion between Kieran Chandler and Joannes Vermorel revolves around the potential of using large external data sources, such as social media, to forecast trends and capture early signals for supply chain optimization. Vermorel argues that, in practice, this approach does not work effectively for supply chains due to the difficulty in extracting relevant data and discerning between real and fake information. Even advertisers struggle to differentiate between real and bot-generated traffic, making it a challenge to gain valuable insights from these data sources. Vermorel suggests that predicting trends using social media might work in rare or extreme situations, but the granularity of information needed for supply chain optimization is not generally attainable.
They discussed the challenges and potential of using external data, such as social media, to predict supply chain trends. Vermorel explains that while there may be valuable information on social media, it is difficult to directly match this data with specific products or SKUs. Furthermore, the quality and context of the data are often unclear, as people rarely mention product details like UPC codes in their posts, and the interpretation of the content can be highly subjective.
Vermorel adds that even when the information appears to be obvious, such as a celebrity endorsing a product, the interpretation can vary significantly. People may focus on different aspects of the product, such as the color, shape, or brand. This fuzziness is compounded by the fact that social media posts are often part of a larger conversation, with interactions and responses adding further ambiguity. Sentiment analysis can help, but it remains difficult to extract clear and actionable information from social media data.
When asked about other early signals for supply chain trends, Vermorel suggests that companies should focus on their own internal data. A sharper view of the future can be obtained by analyzing this data, but detecting early signals for out-of-the-ordinary events remains challenging. He explains that even with a good forecasting model, there are limitations due to the need for stability. A supply chain requires a steady plan to execute, and too much responsiveness can lead to costly false positives.
Vermorel acknowledges that improvements can be made in detecting signals earlier, but the progress is often marginal. For example, at Lokad, they are pleased when they can detect a disruption a few days earlier than before. He cautions that in most cases, these early signals are still reactive, as they are based on company data that reflects changes in sales or other factors.
Addressing the idea of having early signals for significant events, like the coronavirus, Vermorel questions the feasibility of detecting such signals reliably. He suggests that while it may be beneficial to have early warning systems for rare, large-scale disruptions, the practicality of developing accurate and reliable forecasting models for these scenarios remains uncertain.
Founder discussed extreme tail events like COVID-19, the challenges of predicting such events, and the importance of probabilistic forecasting. He argues that predicting outliers is inherently difficult due to their surprising and unprecedented nature. Instead, he recommends adopting a probabilistic approach that assigns a small probability to extreme events and preparing for them over time.
Vermorel emphasizes the need for organizations to be reflective and adaptive in response to world events. He notes that many industries are slow to react, even when given early warning signals, as they lack the ability to efficiently turn forecasts into actionable decisions. This is often due to the manual and time-consuming nature of decision-making processes within companies.
He stresses the importance of having a fully automated process for transforming forecasts into decisions without human intervention. Without automation, the benefits of early warning signals are often lost due to the inertia of people trying to turn forecasts into actions. In addition, he highlights the issue of confidence in early signals, as they are often noisy and uncertain.
Vermorel suggests that the technology could potentially improve in the future, but the key issue remains in turning forecasts into decisions. He believes that a quantitative supply chain approach, involving full robotization of the data pipeline, is essential for addressing this issue. This allows companies to be more responsive and better prepared for unexpected events.
He discussed the challenges of forecasting and responding to spikes in demand, such as during a pandemic. Vermorel explains that the future is inherently uncertain, and probabilistic forecasting is the best approach to embrace this fuzziness. He emphasizes that companies should focus on how to make better decisions based on probabilities, rather than trying to predict the exact numbers.
Vermorel believes that the key to successfully responding to demand spikes is agility, and companies can achieve this by automating their processes. He states that better forecasting technology may only save a company a week or two compared to traditional methods, but the real advantage comes from eliminating the lag between forecasting and decision-making. By integrating these processes, companies can make quicker and more effective decisions.
When asked about the role of supply chain practitioners, Vermorel suggests that they should focus on automating their processes and responding to changes in demand in a timely manner. He notes that companies often waste significant time, sometimes up to a quarter, adjusting to new realities. By streamlining and automating processes, companies can reduce this lag and improve their agility.
Vermorel also talks about the limitations of using social media data for forecasting. While it may provide some insights, it is not necessarily actionable information for companies. Instead, businesses should prioritize automating their decision-making processes and implementing more advanced forecasting methods to increase their agility in response to demand fluctuations.
Vermorel highlights the importance of embracing uncertainty and focusing on making better decisions based on probabilities. By automating processes and eliminating the lag between forecasting and decision-making, companies can improve their agility and better respond to spikes in demand.
Kieran Chandler: Today on Lokad TV, we’re going to discuss whether, with the rise of social media, this is now possible and also what companies can do to take advantage of those early signals. So, Joannes, today our topic’s all about forecasting those early signals. What’s the idea behind this?
Joannes Vermorel: The idea, as it is presented by many software vendors, is that by leveraging large data sources that extend beyond what the company has, you can basically know the future a bit earlier to have a glimpse into a more distant future just a bit earlier. The archetype of this idea of early signal would be to detect that something is trending on Instagram because there are some celebrities who are now wearing a new type of garment and then bam, you can use that to anticipate a month ahead that there is an oncoming trend or something.
Kieran Chandler: So when we talk about those data sources, we’re talking about maybe social media sources. Is that the right sort of data to be looking at?
Joannes Vermorel: At least that’s what many AI-driven supply chain optimization vendors would say.
Kieran Chandler: And in practice, how well does this actually work then, forecasting early signals based on external data sources?
Joannes Vermorel: At present time, there is basically nothing that even remotely closely works for supply chains, not even close. The idea that you can go into social media and just extract relevant data there is just an illusion. It’s people who are pretending to do that, they have never seen the fact that the bulk of the traffic on those social media are not even done by people but by bots, you know, robots. So, it’s very hard to figure out what’s real and what’s not real. Even advertisers, who pay for prints, have a very difficult time sorting out what is real from what is not real, and they are paying a boatload of money to push ads on those platforms. Even when you pay for every single impression and are desperate in making sure that it’s actually real people with real eyeballs looking at those ads, it’s very hard. So, imagine that if you’re just a vendor who is collecting terabytes of random traffic data, what you get is just a lot of randomness, a lot of noise. The idea that you can automatically forecast trends based on that, maybe in some extreme situations, yes, but they are rare, so they are not very interesting. And the sort of information that you would get would not nearly have the granularity that you’re interested in. Yes, if your granularity is predicting who is going to be among the top 50 most popular artists in the U.S. this year, you can have an early signal by just looking at social media. That will give you a clear picture ahead of time of who, at the end of the year, is going to be the biggest artist. But if you actually try to get information such as…As how many t-shirts of this size, this color, this pattern am I going to sell in this region of the US? That’s a very different challenge. Here, I would say mostly you’re not going to have information that is nowhere near the quality of information it takes to translate what you get from those external data sources into something that would make sense supply chain-wise for producing goods and distributing them.
Kieran Chandler: Okay, and so what’s the issue in terms of the quality of data? You mentioned there are lots of robots and stuff on these social media platforms, but still, we’re going to see those posts with the most likes rising to the top, becoming the most visible. So, in terms of data, why is that not interesting?
Joannes Vermorel: The problem is that, again, in supply chains, your productions are organized by SKUs. You’re producing SKUs of something, and then you’re selling those products maybe through a bundle for complex marketing channels, etc. No matter which tweets get retweeted a million times or which Instagram picture gets looked at a bazillion times, there is no direct match between that and any of the SKUs that you produce or distribute. People very rarely include the UPC code in their tweet, saying this product with this barcode is really top-notch. So usually, it’s very fuzzy. Even when it kind of looks obvious, even if you have a superstar that tweets, “This new pair of shoes is super cool,” some people might actually look at the shoes and just interpret the color, some people might think about the shape, others would think about the brand, and some might interpret that it’s anything that looks like that. It can be very fuzzy, and when you compound that with the fact that there is a lot of interaction, so it’s not like you can just look at one post and say, “This is it.” It’s a series; there is interaction. It’s intended for humans. The quality of the response matters because what if someone posts something that has a bazillion retweets, but the retweets are super critical or sarcastic? You can have sentiment analysis, but you’re entering a realm where it’s very difficult.
Kieran Chandler: Okay, so it certainly sounds challenging from the social media perspective, but are there any other early signals that we could look at? We spoke about weather on a previous episode, so what else could we look at?
Joannes Vermorel: Yeah, I mean, usually, if you just look at the data of the companies themselves, there is a lot of data there, and you can try by looking at the data that companies already have to have a sharper view of the future. That’s typically the approach that Lokad does. Then, when it comes to having a true early signal for something that is, I would say, a tail event or something that is a bit out of the ordinary, it’s very difficult. Yes, if you have a very good forecasting model, you can have a signal that comes a bit earlier, but there are strong limits to that. Just because a better forecasting model usually means that…
Kieran Chandler: So, in your experience, stability is important when it comes to forecasting models, be it classic or probabilistic. Supply chains need a relatively steady plan to execute, and you don’t want just a bit of noise in the market to shift gears, which would disrupt production and distribution. How do you deal with forecasting engines that become overly responsive and result in false positives?
Joannes Vermorel: A good forecasting technology won’t go crazy just because of one additional day of the year. For most situations, it’s just business as usual with a minimal course inflection. If you have a forecasting engine that becomes too responsive, you end up with false positives. You might think you have an early signal to change course dramatically, but two weeks later, you realize it was just an artifact or a bump, which can be very costly. We’ve made progress at Lokad over the last decade, but it’s very marginal. We are happy when we can detect a disruption a few days earlier, but shaving off just a few days is the best we can do.
Kieran Chandler: And what about when there’s a large bump, like what happened with the coronavirus? Surely it’s good to have an early signal for these types of scenarios, and what can you do if you have one of those early signals? Isn’t that beneficial?
Joannes Vermorel: I challenge the idea that you can have an early signal for extreme tail events like COVID-19. Most people were puzzled by the way events unfolded, as it was very haphazard and inconsistent across countries or even regions. Outliers, by definition, are very hard to predict. I also challenge the idea that a forecast, even with a two-week head start, would have really helped in dealing with COVID-19.
Kieran Chandler: Probabilistic vision with regard to tail events is the idea that you can’t really predict them. It’s pretty much by design, as it’s very difficult because it’s going to be new and surprising. But what you can do is have your forecast, your probabilistic forecast, that always puts some probability on things that are just wild. It’s not as if you could anticipate COVID-19, but it’s as if you were saying, well, next quarter there is like a one percent chance that business is going to be half of what it is right now or half of what it was last year. Why? I don’t know, I don’t know. It’s just a possibility. It could be a war, it could be a pandemic, it could be a massive recall, maybe a smear campaign online against the brand, or whatever. Is it reasonable to say that for most companies, you have like a one percent chance for every quarter that sales get halved during the next quarter?
Joannes Vermorel: I would say for most companies, yes. It means that four times per century, a massive disaster happens. I think it’s fair. If we look at the 20th century, it was more than that. So, you need probabilistic forecasting, where you put a fixed, low probability on extreme events, and then you prepare yourself against that, knowing that this can happen anytime, and when it will happen, you will assume that you will have no warning. But the good news is that if you have this probability in place for years, it means that you’ve done small adjustments that, if this thing happened, you’re kind of prepared. But it took years of preparation, making good use of your resources.
Kieran Chandler: We always speak about this idea of being reflective to what’s going on in the world. Are there any industries that you’ve seen that are classically slow in their reactions that could benefit from maybe having an early signal approach?
Joannes Vermorel: The thing is that most of the industries that I’ve observed, even if you give them an early signal, they would not do anything with it. The reason being is that until you’ve reached a point, and that brings the question back to this quantitative supply chain vision, where the point number four is about having full robotization of your data pipeline so that you can generate your decisions. You see, what happens in most companies is that the forecasting element is just one step of the process that ultimately leads to decisions being taken, the decision being things like, should we buy more from suppliers, should we produce more, or should we actually move the price up or down? But the reality is that for most companies, the process to turn forecasts into actual decisions is very manual and, in my perspective, very slow. So, imagine you’ve invested a lot in technology to get an early signal, maybe one week earlier, two weeks earlier if you’re very good. That’s the sort of thing that you can get with technology as it exists right now.
Kieran Chandler: Now, there might be edge cases where you can see something in advance, but I doubt in this case that it’s a statistical forecast. It’s more likely going to be high-level insight about the evolution of the market, driven by numbers and statistics. In this sort of situation, you have your early forecast, but are you going to act on it?
Joannes Vermorel: Usually, if the consequence of having an early signal is just sending an email to someone in your company or expecting that person to go to an app to look at the alerts, nothing will happen in the short amount of time. It’s going to be sluggish. Supply chains are complex and distributed, involving many people, systems, and machines across many locations. If you want to have a good supply chain response, you need coordination. Unless there’s a super high degree of automation in place, that coordination will involve real people making frantic phone calls, and it’s going to take time, typically weeks.
The bottom line is that if you have an early signal but don’t have a fully automated process that turns your forecast into decisions automatically, without human intervention, the vast majority of the potential gain from the early signal will be lost in the inertia of people trying to turn those forecasts into actions.
Kieran Chandler: I guess one of the problems with these early signals is the amount of confidence you can actually have in the results. Like you said, they’re often fairly noisy. Do you think in the future we could get to a stage where the technology is in place that you can have more confidence in the results?
Joannes Vermorel: No, and the reason is that the future is fuzzy, and this uncertainty is irreducible. What we’ve been focusing on for the last decade is probabilistic forecasting, which embraces the fact that the future is uncertain. If you expect a magical computer system that will give you the winning lottery number with 100% certainty, it doesn’t work like that. At best, you can have a system that gives you probabilities.
For example, you might be facing a situation that starts as having a 1% chance of being an extreme event. Then, maybe you start seeing something off, and the next day, the estimate increases to 1.5%. Should you do something? Maybe not. The next day, it increases to a 3% chance, which is an exponential growth, but still very slow. Then it’s 6%. Should you start going crazy because there’s a 6% chance?
Kieran Chandler: To happen, like, you know, we are talking of something that was only supposed to happen once per 25 years, and here we are, something that would happen maybe once every five years or so. But it’s, again, this is just a bump, and the idea is that those priorities are very fuzzy, and if your decision system is well-engineered, the things that transform those priorities into the decision will basically gradually reflect that the risk that was at one percent is now six percent. So it will steer a little, but just a little, the decisions into a direction that make it more protective for your company, but without going crazy again. You don’t want to have something that just drives your supply chain crazy by making wild moves in changes, and maybe the next day, things will go back to three percent estimate, you know, from six to three because actually it was maybe a false positive or maybe it will keep increasing. But you see, that’s a sort of good response that you can have from a fully automated data pipeline. And then, if you have supply chain scientists on top, and they are seeing the news, they have a high level understanding, and they see that something like a big wave is coming, like a coronavirus, then you can actually have people that will, on top of that, kind of twist the numerical recipe itself so that it takes a bit of extra human insight to steer the system into a better direction.
Joannes Vermorel: So, if you’re a supply chain practitioner who’s watching this, should their main focus be on fully automating their processes instead of worrying about what happens maybe earlier on in the process? They should be looking at maybe how they respond to it in the moment.
Kieran Chandler: Exactly. I mean, you have to assess how many days of agility you are going to win. For most companies, I would say with a better forecasting technology, maybe like the one of Lokad, let’s say, I mean, how much can you save? Maybe a week, maybe two, compared to a moving average. Maybe you have better forecasts, so it might count as a little bit better, you know, maybe four weeks if your company was very bad at dealing with seasonality or something. But, I mean, if you didn’t have a dysfunctional process in place, probably we are just talking of a few weeks. But when it comes to decisions, I frequently observe that companies can waste up to a quarter to get their act together and accept a new reality. So it’s very slow, and even in super reactive businesses like fast fashion, where it was taking six weeks to transform the forecast into a decision, and I’m talking about a decision and then you still have to produce, transport, and distribute. It still takes time. So really, I think on average, for the vast majority of companies, it’s a much safer bet to say we are just going to first eliminate the lag that is between decisions, between forecasting and decisions, by just putting all those things together into one automated process and then figure out how they can have super advanced forecasts where they can shave a few more days or weeks if they are lucky. Okay, great, we’ll have to leave it there. But I guess now we can use social media without knowing that companies are watching everything we do.
Joannes Vermorel: Oh, they’re watching. It’s just nice, so it just makes some analysts happy, but they don’t turn it into anything actionable.
Kieran Chandler: Great, so that’s everything for this week. Thanks very much for tuning in, and we’ll see you again in the next episode.