00:00:04 Impact of promotions on forecasts.
00:00:51 Promotions variability across industries.
00:02:12 Challenges in predicting promotion-driven demand.
00:03:58 Analysis: Traditional supply chain methods’ risks.
00:06:00 Critique: Moving averages and forecasting methods.
00:08:01 Criticism: Simple promotion forecasting models.
00:08:36 Companies’ reliance on moving averages.
00:10:26 Data needs for promotional forecasting.
00:13:25 Machine learning’s role in promotion forecasting.
00:15:09 Promotional data’s role in non-promotional forecasting.
00:17:34 Machine learning systems’ learning speed.
00:20:09 Difficulties and strategies in promotion forecasting.
00:20:35 Implementing promotion forecasting in organizations.
00:23:17 Strategic execution of promotions using machine learning.
The interview between Kieran Chandler and Joannes Vermorel, the founder of Lokad, is revolving around promotions, forecasting, and supply chain optimization. Vermorel is explaining that promotions can distort actual demand, thereby posing a challenge in demand forecasting. Many companies are resolving this by revising sales history to neutralize promotional impacts, but this approach risks creating misleading forecasts. Companies are mostly employing moving average methods for supply chain optimization, which are proving insufficient in anticipating promotional uplifts. Improved forecasts are requiring more complexity, like machine learning, but also demanding high-quality data on promotional strategies. Vermorel is highlighting the long-term influence of promotions and the need for strategic planning before implementing machine learning in promotion forecasting.
The conversation between Kieran Chandler and Joannes Vermorel, the founder of Lokad, is revolving around the topic of promotions and their implications on forecasting and supply chain optimization.
The discussion begins with an examination of how promotions have a drastic impact on sales, potentially positive if executed correctly and negative if poorly handled. They are also shaping customers’ expectations about discounts, thus affecting buying behaviors. Chandler and Vermorel acknowledge that promotional strategies significantly differ across industries, with certain sectors like luxury goods avoiding discounts, whereas others like hypermarkets are leveraging daily promotions.
The conversation subsequently ventures into the complexity of promotions from a forecasting perspective. Vermorel elaborates that even though promotions are generally linked with an increase in sales, this increase does not necessarily correspond to an increase in actual demand. This discrepancy emerges because observed sales during a promotion may exceed actual demand as consumers seize the opportunity of lower prices, even if they don’t need the product immediately. Such behavior can distort the perception of real demand.
The conversation then pivots to the supply chain implications. Vermorel points out that when planning for future demand, supply chain practitioners have to account for the distortion caused by promotions. However, many companies are relying on traditional supply chain methods, trying to obscure the effect of promotions by rewriting sales history, thereby effectively erasing promotional sales spikes. This approach, while trying to normalize sales data, introduces risks as it substitutes real historical data with artificial constructs, which can lead to misleading forecasts.
Chandler is expressing concern about these potential dangers, prompting Vermorel to delve into why companies choose this method. He details that most companies deploy technologies primarily based on variations of moving averages for their supply chain optimization. These methods, although somewhat adequate for non-promotional periods, fail to foresee promotional uplifts, leading to subpar forecasts. This performance gap is what prompts many of Lokad’s clients to pursue improved promotional forecasts.
Vermorel starts by recognizing the significant role of statistical models in enhancing forecast accuracy, noting that while they are effective, they are not the single solution to accurate forecasting. He discusses the widespread use of moving averages in the industry due to their simplicity and accessibility, especially given that a large proportion of supply chains still operate using basic tools like Microsoft Excel. However, he indicates that while moving averages are simple, obtaining better results requires significantly more complexity.
The conversation then moves to the concept of machine learning, which presents a substantial advancement in complexity and potential results compared to moving averages. Machine learning demands not only a more sophisticated understanding of statistical methods, but also a higher standard of data quality. Particularly, Vermorel underscores the necessity for comprehensive, high-quality data about promotional strategies.
Companies need to collect extensive data on all aspects contributing to the mechanisms of promotions. This includes not only price adjustments but also the marketing efforts accompanying promotions and the visibility strategies employed. Vermorel examines various industry-specific examples, such as end-of-season sales in fashion or product placement strategies in retail markets.
Vermorel also stresses the importance of data quality. He explains that while minor inaccuracies or gaps in data may not instantly cause a company to go bankrupt, they can severely hamper the performance of machine learning algorithms that rely on accurate data. He suggests that achieving robust historical promotional data typically demands considerable effort over several months.
The interview then turns towards the evolution of data collection and its significance in promotional forecasting. Vermorel proposes that with accurate data, forecasts can more effectively predict which promotions to use and when they might work best. Machine learning algorithms process the data by identifying and flagging periods where sales were either inflated or likely suppressed due to promotional activities. This strategy, Vermorel explains, helps improve the accuracy of forecasts even outside of promotional periods.
At first glance, this concept appears counterintuitive. However, Vermorel justifies it by explaining that promotions have a ripple effect on other, non-promoted products. For example, a significant promotion on one product can cannibalize the sales of competing products, thereby making forecasting more complex. Understanding these indirect effects of promotions thus adds another layer to supply chain forecasting.
Vermorel emphasizes the necessity to anticipate future events and their potential impact on the supply chain. He proposes that if businesses could communicate such future events to a machine learning-powered forecasting engine, it could leverage that data and incorporate the upcoming event into its predictions.
Chandler asks about the learning capacity of machine learning and how quickly it can yield results. Vermorel clarifies that it’s predominantly a statistics game. The speed at which the machine learning algorithm learns is dependent on the frequency of promotions. For instance, if there’s only one promotion per year for a product, learning would be slow. However, the algorithm learns faster with regular promotions as it can use data from similar past promotions.
Additionally, Vermorel clarifies that the prediction of promotions is not about forecasting a single time series but about examining the typical impact of a promotion under similar conditions, like discounts or communication channels. He provides examples from the fashion industry, where end-of-season promotions are common, and e-commerce, where products are continually promoted on the website’s front page.
Addressing CEOs or aspiring CEOs, Vermorel outlines a process to introduce promotion forecasting within their organizations. He underscores the need for data collection, emphasizing the importance of specific data over broad datasets. He recommends collecting detailed data about the promotions themselves: the products, the promotional mechanisms, and other variables such as free shipping.
Vermorel highlights the necessity of having a quality assurance process to ensure the accuracy and relevance of data. He also encourages leaders to ponder about the underlying purpose or endgame of their promotion strategy, as it differs across various industries. He provides examples from the fashion and general merchandise industries, each having unique objectives behind their promotions.
He urges organizations to consider the long-term impacts of their promotions. These activities, he states, educate customers in a certain way, which can have enduring effects. Therefore, businesses must strategically think about the kind of influence they want to exert over their customers through promotions.
Machine learning comes into play once organizations have clarified their strategic thinking and gathered the relevant data. Vermorel reiterates that machine learning, despite being mechanically intelligent, will not devise high-level strategies, emphasizing that it is essential for businesses to strategize before employing machine learning for promotion forecasting.
Kieran Chandler: Today on Lokad TV. We’re going to be talking about promotions and the impact that they can have on forecasts. Promotions can be incredibly variable and difficult to predict. Do them right, and they can vastly increase your sales, but do them wrong and they can reduce the credibility of your products and alienate your customer base into expecting vast reductions in price. Promotions often hit headlines for all sorts of the wrong reasons. During Black Friday, customers can be seen fighting over priced products in department stores. However, these periods are undoubtedly important for the retailers who often report massive increases in products. So Joannes, promotions are incredibly variable depending on the industry you’re in. To kick things off, perhaps you could sort of explain the companies that we’re talking about here.
Joannes Vermorel: Yes, so indeed promotions come in a variety of flavors depending on your vertical. It runs from industries like luxury, that never do any promotion. I mean, you will not get a Rolex on sale, to industries that have promotions on a daily basis, such as hypermarkets, where you would see daily promotions on things. Promotions can be about price, that’s what you typically get in a hypermarket where you buy two, you get one for free. But promotions also in e-commerce are about putting something forward. If you send a newsletter to half a million subscribers to put forward a product, you will see a big increase in sales, even if you’ve not actually decreased the price of the product by even 1%. So indeed, it’s the same word, but it flags very different realities depending on the company you’re talking to.
Kieran Chandler: Okay, so they’re generally going to be a pretty good thing because we’re seeing normally an increase in sales. Why do they complicate things? Why is there a difficulty seen with promotions?
Joannes Vermorel: They complicate things because what you want to predict, if you want to optimize your supply chain, is future demand. The problem is that you do not observe future demand or even past demand, you observe past sales. The sales are not the demand. The sales come with all sorts of bias. If you don’t have any stock left, you will have a stock out, so you have like zero sales, but demand is still there. Promotions are just the opposite effect, in that you can see a lot more sales than you have actual demand. For example, if you have a massive price drop, even if people don’t need it right now, they would maybe start to build up their own stock at home with the product you’re selling, just because they feel that it’s a good opportunity. So the sales can give you a misguided sense of what the demand actually is. Most supply chain practitioners know that they should not just apply a moving average on top of promotions because the promotion is pushing the observed demand naively to the roof. You know that at the end of the promotion, sales will drop. That’s the very basic phenomenon that you need to take into account; otherwise, your planning is going to be completely wrong.
Kieran Chandler: So you’re saying there’s a real misguided sense of demand. How are companies taking that into reality? How are they adjusting towards that?
Joannes Vermorel: The traditional supply chain method consists of rewriting the sales history to mask the effect of promotions. Most companies will take their sales history, look at the promotional period, see the spike, and try to erase spikes from their sales history. That way, when they apply something that is very akin to a moving average, the moving average is not overly biased by the promotion that just took place. I’m not saying that’s the right thing, but that’s what most companies are still doing nowadays.
Kieran Chandler: And removing the spikes, which is seen by promotions, sounds very dangerous. Aren’t you going to get sort of two versions of the truth? Why are they actually doing this?
Joannes Vermorel: You’re completely spot on; it is very dangerous. The reason is that you’re replacing historical sales data, which is accurate, with mock data that you just made up. It’s a distortion of reality and then you’re going to base your planning on this made-up data. Yes, there is a real danger. It’s a methodological danger. If your corrections are misguided, then your forecasts will be built upon flawed inputs, which could exacerbate the problem in terms of supply chain planning.
So, why are people doing it at all? If you look at the sort of technologies that most companies are still using for their supply chain optimization, it’s basically a glorified version of moving averages. Moving averages go by different names - exponential smoothing, Holt-Winters - but essentially they’re all variations of moving averages. These methods are just a bit more nuanced to deal with seasonality, but the crux of it is still a moving average.
This is why if all you have is a moving average, you need to make your historical sales compatible with it, which is the only mathematical model that your company has available. But it’s a flawed perspective because there’s a lot more than just moving averages.
Kieran Chandler: But these statistical methods are surely what a lot of these companies are basing their future decisions on. So, do they actually work? They must be good enough if companies are using them all over the world.
Joannes Vermorel: The reality is that they are not really good enough. Most of the companies who became our clients were actually telling us that one of the key motivations was to get better promotional forecasts because it was a real problem for them.
Your moving average can cope, but it cannot anticipate the uplift or the effect of the promotion. While it may not completely ruin the forecast for non-promotional periods due to the promotion, it still doesn’t provide anything to actually cope with an upcoming promotion. So, the moving average saves you a little bit, but it doesn’t address the problem at its core. It doesn’t produce an actual promotion forecast, it just prevents the promotions from skewing all the non-promotional forecasts. And by the way, this method is very weak; it doesn’t even work correctly and it comes with a lot of problems.
Kieran Chandler: So why are these companies so happy to work with these moving averages? Why are they not changing? Why is it still sort of a problem that exists?
Joannes Vermorel: I believe that a moving average is something that any engineer can think of in like two hours. So simplicity is very powerful. You will end up reinventing in two hours a moving average and some sort of recipe to actually make your moving average work, which is a very powerful force.
Remember, about 80% of the worldwide supply chains are still operating on top of Microsoft Excel, so a moving average on Excel is very, very easy to write. That’s probably the reason.
Another reason is that if you want to do better, it’s significantly harder. Suddenly you have to jump from moving average to machine learning. So, you have to go from a very simplistic method that anyone can understand, to machine learning which is a lot more complicated.
Kieran Chandler: We’re discussing an algorithm that can work within two hours, even if you’ve never done machine learning before. But to implement it successfully, you also need to pay close attention to the quality of data about your promotions. Could you describe how this process works, and how we can proceed with better methods?
Joannes Vermorel: Of course. Regardless of any alternative to moving averages, we have to rely significantly on data. This factor makes the game a lot more complex.
Kieran Chandler: If we’re aiming to do things better, what sort of data should we be collecting? Should it be marketing efforts, price changes, or something else?
Joannes Vermorel: The company needs to collect everything that contributes to the promotional mechanics. The price is one thing, yes. If you discount the price of a product but don’t inform anybody about the discount, nobody notices, except the people who were already intending to buy the product.
So, a promotion is not just about price. It’s also about spreading the word. Letting the market know you have a promotion is crucial. In certain industries like fashion, there are end-of-season sales. Everyone expects these sales, which are a specific type of promotion. But in some domains, it’s a completely different game.
For example, in hypermarkets, it’s not just about having a price drop. Often, it’s about moving a product to the head of the gondola—the super-premium placement at the end of the aisle where products are highly visible. Even better, you could place a large pile of the promoted products at the entrance of the store.
So, now the question becomes: does your ERP system properly track all this data? If you fail to properly keep track of what you sell or buy, your company could go bankrupt. If you don’t know what you’re selling or buying, either customers or suppliers could defraud you, which would lead to bankruptcy.
On the other hand, if the date of your promotion is recorded incorrectly in your ERP records, it won’t bankrupt your company. But if you want to feed a machine learning algorithm that forecasts promotions with inaccurate data, it won’t work.
So, you need a quality assurance process for your promotional data. From our experience at Lokad, this process can be a lot of effort. For most companies, it requires months of effort to obtain quality historical promotional data.
Kieran Chandler: So what you’re saying is there’s been a real evolution in the industry because it’s not a critical metric to measure, but an evolution of how we measure it?
Joannes Vermorel: Yes, it’s very difficult to measure.
Kieran Chandler: Hypothetically, if a company has measured all these things and they’ve gathered all that information, does that mean through your forecast, you’ll be able to tell us what promotions to do, when they work best, and what sort of information you would be able to generate?
Joannes Vermorel: Yes. The first thing to understand is that instead of correcting historical data, a machine learning algorithm looks at it from the perspective of enriched historical data. You don’t try to tweak your sales. You’re going to flag the periods when the product sales were inflated and the periods where they were probably censored.
For instance, at the end of a promotion, if people have built up their own stockpile of goods, you might see a drop in demand. You might even have a period of time where you will observe some kind of censored demand. So, you need to account for all these factors.
Kieran Chandler: So, the first benefit of promotions, which may not be immediately apparent, is the potential to improve the quality of your forecast even during non-promotional periods. This is achieved through the use of an algorithm that is better at predicting biases. What you’re saying is that promotional data is used even when promotions are not occurring, correct?
Joannes Vermorel: Yes, it might seem counterintuitive, but let’s consider an example. If you have a major promotion for a variety of tomato, it will likely cannibalize the sales for all other non-promoted products. It’s quite obvious that a significant promotion for a particular product will have a massive cannibalization effect on all other products that are in competition with the promoted item. Hence, promotions make forecasting more complex, not only for the product being promoted but also for all non-promoted products.
Kieran Chandler: I see, so it’s not as counterintuitive as it first appears, but it requires some consideration. Now, let’s move on to the topic of anticipating future promotions. If you are aware of an upcoming promotion, you can anticipate the uplift, or the increase in sales, that will result. However, this raises the question of your process for deciding about future promotions. Could you elaborate on how these decisions are made and how they are inputted into your forecasting engine?
Joannes Vermorel: Sure. The first part of the process is deciding on future promotions. The second part involves ensuring that these decisions are inputted into our forecasting engine. This isn’t just about past data, but also about future expectations. If we’re planning an event that will significantly impact the supply chain, the forecasting engine needs to know about it in advance. A machine learning-driven forecast engine, if informed in time, will be able to adapt and reflect the upcoming event in its forecast.
Kieran Chandler: That’s fascinating, especially the idea of a promotion of the decade. But considering the complexity of machine learning, how quickly can these machines learn and adapt? When can one start expecting to see results?
Joannes Vermorel: It largely depends on the number of promotions. It’s a game of statistics. If you’re running a promotion only once a year for a product, it’s hard to learn anything. Remember, forecasting promotions isn’t about forecasting a time series. Each product might only get promoted once or twice in its lifetime. If you want to understand the impact of a promotion, you need to consider what the typical impact of a promotion is under similar conditions—price discount, promotional mechanics, communication channels, and so on. For instance, in the fashion industry, which does end-of-season promotions, you have several data points since it happens four to eight times a year. For a hypermarket, every single product is potentially a data point.
They have hundreds of products that are being promoted each day, and they rotate, it’s not always the same products. It’s something that happens all the time in e-commerce. Typically, you always have one or two products that happen to be on the front page of your e-commerce website, so there’s a strong promotional mechanism that is happening all the time. But it gets very difficult when you start thinking about forecasting promotions. A question our customers ask is if our forecasting engine can maybe forecast promotions, but what they really would like is to decide what is the best promotion, which is a completely different question and a very difficult one.
Kieran Chandler: Let’s say some of our viewers might be CEOs. What’s the process they should follow if they want to actually implement promotion forecasting into their organization?
Joannes Vermorel: First, they need to collect all the relevant data. I’m not talking about collecting Twitter data or what people are saying on Facebook. It’s more about knowing what is the list of products that get promoted, what is the exact promotional mechanism. They probably have their own categories of buy one get one free, or is it a price drop in percentage, or any kind of promotion with free shipping for example. The first step is to collect what I call quasi transactional data. It’s not transactional data because it doesn’t appear in the invoices or payments, but it’s very well specified. You need to collect them, make them part of your system, and have a quality assurance process on top of that.
The second step is to really think about why you do promotions, what is the endgame. The problem is that it’s typically a very different endgame depending on your vertical. For example, in fashion, the goal is to liquidate all your old inventory so that you can always sell stuff that’s up to date with the latest trend. In general merchandise, promotions are typically not triggered by the retailer but by their supplier, as a negotiation to increase awareness for a new product, like a new flavor of a fast-moving consumer good. The end games are very different, which means that when you want to assess how you should execute, you need to think about the impact of your promotions.
You are trying to influence your customers, and you should really think about exactly what kind of influence you want to create. For example, if you educate your customers to always buy your stuff cheap because they know a promotion is always coming, then your customers will learn that they just need to be a bit patient and they will get it for cheap. It requires really strategic thinking and this strategic thinking should happen before the machine learning.
Once you get this right, you can have a machine learning system that is smart and capable of processing all the data that you collected and aligned with your strategic thinking. Machine learning is very mechanical, so it will not be able to execute any high-level strategy, that’s your job.
Kieran Chandler: I’m afraid we’re going to have to wrap things up for today. When you talk about educating customers, I’d probably count myself as one of the least educated customers there are. That’s all for this week, thank you for watching and we’ll see you again next time. Goodbye.