00:00:03 Significance of historical data in Lokad’s tech.
00:01:40 Depth of history for supply chain optimization.
00:04:24 Key datasets: catalog, orders, sales, current stock.
00:04:56 Catalog data’s role in forecasting, seasonality.
00:07:06 Comprehensive sales history data’s importance.
00:08:42 Tracking product’s sales cycle nuances.
00:11:16 Purchase order history in price prediction.
00:13:53 Profit margin, inventory risks, forecasting difficulties.
00:14:32 Accurate recording of order dates, deliveries, QC impact.
00:16:01 QC in strawberry shipments, order influence.
00:16:55 Understanding current stock levels’ significance.
00:17:36 Theoretical stock levels can cause inaccuracies.
00:18:39 ‘Stock outs’ and historical stock level importance.
00:20:59 Secondary datasets: promotional activity.
00:22:00 Promotional activities’ impact on sales trends.
00:23:13 Common mistake: Complex ERP system choice.
00:24:39 Mistake two: Gaps in order data recording.
00:25:35 Mistake three: Neglecting stock level snapshots.
00:26:30 Importance of accurate data recording.
Summary
In an interview with host Kieran Chandler, Joannes Vermorel, the founder of Lokad, outlines the complexities of supply chain optimization. He highlights the importance of the historical depth and detail of data, a challenge that many companies grapple with. Vermorel suggests that ten times the lead times’ worth of historical data is needed for accurate forecasting. He emphasizes the necessity of four key datasets: catalog, purchase order history, sales history, and current stock levels. According to him, these datasets help in revealing patterns, optimizing forecasts, and managing stock levels effectively. Finally, Vermorel identifies common mistakes made by clients, such as opting for difficult ERP systems and failing to record purchase orders or historical stock levels.
Extended Summary
In the interview, Kieran Chandler, the host, is engaging in a conversation with Joannes Vermorel, the founder of Lokad, a company that specializes in supply chain optimization. They are discussing the complex nature of data requirements for Lokad’s technology and the challenges customers encounter when trying to adapt to these requirements.
Vermorel is explaining that the primary challenge is the historical depth and detail of data necessary for quantitative supply chain optimization. While the current state of a company is relatively easy to grasp, comprehending the company’s history necessitates analyzing a more extensive set of data. Numerous companies encounter difficulties when attempting to collect the relevant data for this, particularly if they have not previously concentrated on such details.
The extent of history needed, Vermorel suggests, varies depending on the type of business. A historical data of less than a year makes it hard for Lokad to detect any seasonality, and this would be the case for any statistical method. It starts to work effectively with three years of data. Nevertheless, the quantity of historical data also relies on a company’s lead times. A company with short lead times can function with less data, while those with longer lead times necessitate a more profound depth. As a general guideline, ten times the value of the applicable lead times’ history is suggested for accurate forecasting.
Vermorel is stressing that there is no magic in forecasting: without proper data, neither statistical forecasts nor manual ones will be precise.
The conversation then shifts to the four main datasets that Lokad employs: the catalog, the purchase order history, the sales history, and a snapshot of current stock levels. Vermorel is expounding on the importance of the catalog, which enables their forecasting technology to leverage the relationship between products to refine the forecast for each. This proves particularly beneficial when forecasting new products, which have no sales history. Attributes that describe the product, like product categories, sizes, and colors, can be utilized to deduce seasonal patterns based on similar products.
Sales history is essential, Vermorel is stating, but it’s critical to verify its completeness. An incomplete sales history can occur from missing sales data from products no longer sold. At this point, the conversation is interrupted, with Vermorel promising to provide more information about the significance of sales history.
Vermorel is commencing by underscoring the necessity for complete sales history, inclusive of details on products no longer in sale. This is because in certain instances, these products are eliminated from the catalog, resulting in skewed sales history data. He also mentions the difference between sales and demand, claiming that sales do not always mirror actual demand due to varying factors. For instance, the rate of sales immediately following a product launch can provide an estimation of future demand.
Vermorel then talks about the importance of accurately recording the start and end dates of product sales. This data can reveal patterns in consumer behavior, like sudden shifts to new products or abrupt discontinuation of sales due to removal from an online catalog. It’s also important to track the sales history across different channels, like physical stores and different online platforms.
The conversation turns to purchase order history. Vermorel is explaining that it’s beneficial to observe how much has been paid for various stock items over time. However, he points out that predicting future prices is complicated due to factors such as currency fluctuations. Although these fluctuations can influence margins and inventory decisions, Vermorel asserts that forecasting these changes isn’t the primary focus of Lokad.
Vermorel also emphasizes the importance of understanding the dates associated with purchase orders and deliveries. These dates provide insight into lead times, which are critical for supply chain optimization. Lead times can vary seasonally or due to events like Chinese New Year. He adds that it’s essential to document any discrepancies
between ordered and delivered quantities, as this can affect future orders, especially when a fraction of goods don’t pass quality control.
Vermorel is highlighting the necessity of having a current snapshot of stock levels. He acknowledges that in theory, these levels could be reconstructed from sales and purchase order histories, but inventory inaccuracies often arise. These inaccuracies can accumulate over time, leading to significant discrepancies. Therefore, maintaining regular inventory snapshots and controls is crucial for effective supply chain management.
Vermorel is discussing the significance of understanding current and historical stock levels. The stock level informs the need to reorder but also offers insights into sales history. For instance, stock outs—instances where the demand couldn’t be met due to inadequate stock—shouldn’t be interpreted as a lack of demand. In reality, unmet demand due to stock outs is still demand, albeit unobserved. Historical stock levels, thus, help determine the frequency and impact of stock outs. Vermorel also considers a more nuanced understanding of stock outs, where it’s not always about having zero units in stock. In certain cases, if customers require a specific quantity of a product, not having enough to meet that quantity could also be considered a stock out.
The conversation is then moving on to the importance of secondary data sets. Vermorel is indicating that promotional activities are a crucial data set to consider, not just in terms of pricing but also about what products are being promoted through various channels. This data can explain spikes in the sales history. For instance, a spike could be due to a promotion or a genuine increase in demand. Understanding this distinction assists in forecasting future demand.
Vermorel also touches upon common mistakes made by clients. The first is selecting an ERP system from which it’s tough to extract data. This choice can lead to a significant struggle in data extraction, even if the system is local. The second mistake is not recording purchase orders. While small companies may manage their orders with spreadsheets, this approach becomes problematic when trying to analyze lead times. The third mistake is not recording historical stock levels. Given the low cost of storage, Vermorel argues that there’s no excuse not to keep this data.
Full Transcript
Kieran Chandler: Today on Lokad TV, we’re going to investigate the exact depth of data that is required to work with us and also understand some of the real shortcomings that some of our customers have faced. So, Joannes, why is it that customers aren’t quite able to work with the Lokad technology?
Joannes Vermorel: One of the biggest challenges customers face is indeed data. In practice, it should not be, but it usually is. The key insight is that the data you need to operate your company is not the same as the data you need for quantitative supply chain optimization. For the latter, you need to have the entire history of your company. Not going back two decades, but you certainly need a certain depth and a certain level of detail in your history. Many companies that have not succeeded at setting up such a system in the past, or where it hasn’t even been tried, haven’t necessarily paid attention to all the details to have all the relevant data. Thus, they might face complications when it comes to gathering all the relevant data to get really relevant results.
Kieran Chandler: We talked about the depth of history. What is the minimum timeframe of depth of history you should have?
Joannes Vermorel: The depth of history depends on the type of data and the type of business that you have. As a rule of thumb, if you have less than a year, it’s very hard for Lokad to capture any seasonality. This would be true for any statistical method. If you have less than a year, you have not observed the seasonality of the company. To start looking at seasonality with some degree of accuracy, you need about 18 months. It will start working correctly with two years and will work very well with three years. But the amount of depth of history that you need also depends on your lead times. If your lead time is one day, you can work just fine with just two months of history because you don’t really care about seasonality that much. But if your lead times are four months ahead, then it becomes much more important to capture this seasonality. So, if you want something that works well, you would need to have at least probably something like 10 times your applicable lead times worth of history.
Kieran Chandler: Let’s talk about the data itself now. At Lokad, we work on four key datasets, don’t we? The key datasets are the catalog, the purchase order history, the sales history, and also a snapshot of current stock levels. Let’s look at these in a little bit more detail. If we start off with the catalog itself, why is it of interest? I mean, surely our clients have a good understanding of what it is they’re selling and the exact details on that. What more can be said about this catalog?
Joannes Vermorel: The catalog is of prime importance. It’s a very significant part of the data.
Kieran Chandler: We’re discussing Lokad’s forecasting technology and how it takes advantage of the relationship between products to refine the forecast of every single product. Can you elaborate on this?
Joannes Vermorel: Absolutely, for instance, when forecasting new products, we don’t have any sales history to rely on. So, we rely on the attributes that describe the products. We use these techniques extensively for products that have already been launched.
Kieran Chandler: Can you give us an example?
Joannes Vermorel: Sure, let’s take a product that has been sold for three months. Can we apply a seasonal pattern to the projected demand for this product? The answer is, if you only look at the history of this product, you can’t because you have just three months of data. However, if your business has been around for years, you can look at comparable products and detect their seasonality. So, with only three months of history plus the attributes of the products, we’re able to infer the applicable seasonality for this product.
Kieran Chandler: So, the catalog, with things like product categories, product attributes like size and color, is very important?
Joannes Vermorel: Yes, absolutely. The catalog is essential. We also process the plain text label if it’s present, which can help in domains where the catalog itself is poorly structured. The goal is to extend these attributes so that we can operate all these correlations and refine the forecast for every single product, even if the information we have on this product is limited.
Kieran Chandler: Let’s talk about sales history. From a forecasting perspective, I assume sales history is the most important thing to look at. What sort of information are we interested in other than that?
Joannes Vermorel: The first thing is to make sure that your sales history is complete. It can be incomplete in subtle ways. For example, if you don’t have sales data from products that you don’t sell anymore or if you don’t have the product information associated with those sales data. To have a complete sales history over the last three years, you need to have all the sales that did happen at the time, including products that you’re not selling anymore.
Kieran Chandler: So, old products that aren’t being sold anymore can skew the data?
Joannes Vermorel: Exactly, we sometimes see setups where products that are not being sold anymore are purged from the catalog and sales history. This results in a biased sales history where the data from products that are not being sold anymore are not readily available. Another subtle aspect of sales history is to remember when you did start selling a product. Sales are not the same as demand. There are biases like, if you start selling a product and have zero demand for a month, then finally sell one unit, it’s different from selling a product the same day it goes online.
Kieran Chandler: Can you give more insight on this?
Joannes Vermorel: Sure, if you put a product online and sell one unit the same day, it suggests that you might end up selling one product per day. However, if you put a product online and have to wait one month to sell your first unit, it suggests you will be selling like one unit per month. So, understanding when you started selling a product is also crucial.
Kieran Chandler: Can a statistical engine make the difference? The forecasting engine can only make the difference between two situations if you have properly recorded the date when the demand actually started, and the end date when you stopped selling the product.
Joannes Vermorel: Absolutely. But it’s also important to know why a product stopped selling. Did it drop suddenly due to a new better product entering the market, triggering a sudden shift in consumer behavior? Or did it stop because you removed it from your online catalog or store shelf?
Same thing, you need to have a proper recording of the end dates. There might also be other factors if you have a product that is on and off the market, or if a product is sold but not on all your channels. For instance, if you’re selling in physical stores, online, or maybe through a B2B channel, you need to record the history of what was influencing the demand, such as which channels were available at any point of time.
Kieran Chandler: Let’s move on now to the history of the purchase order. I guess the nice thing about looking at the history of the purchase order is you can see how much you’ve paid for various items of stock. So, I’m assuming we use that to forecast going forward what is the likely price we’re going to pay for an item. Is that correct?
Joannes Vermorel: It depends. Because the likely price depends on the supplier. For most verticals, the price you pay to suppliers is relatively stable. However, there can be surprises, typically due to currency fluctuations. For example, you might end up with a quote in yuan while you’re actually selling stuff in dollars. If there is a 15% fluctuation in currency exchange rates, you can have a surprise where things are significantly cheaper or more expensive.
Kieran Chandler: Is that something we can include in the forecast, that kind of currency fluctuation?
Joannes Vermorel: It depends on what you want to forecast. Currency fluctuations will have an impact, but not directly on your demand forecast. It will have more of an impact on how you optimize your decisions on top of the forecast.
For example, if your margin for a given product is higher or smaller, you would not take as much inventory risk on it. If you have an item that has a massive margin, it’s literally a crime to have a stock out. In an extreme case, if you have a 95 percent margin, selling one unit pays for the overstock of twenty other units. Therefore, you want to be overstocked if you have items that are steadily selling with such a high margin.
On the other hand, if you have a 5 percent margin, you need to be very careful with how much you have in stock. If you have a bit of overstock, you’ll need a huge amount of sales to compensate for the money that’s frozen into the one extra unit in stock. If you end up doing an inventory write-off, it’s going to be very costly.
In general, we don’t typically forecast future margins because it depends on so many factors. We can, but typically it’s not the primary thing because nobody can really forecast currency fluctuation. The only people who can are those directly trading on the market to make money out of the expected currency trading. That’s not our specialty, so we don’t pretend to be better than the market at doing forex trading.
Kieran Chandler: Let’s move away from trading then. But in terms of purchase orders, what else is of interest?
Joannes Vermorel: So typically, the dates involved in a purchase order and delivery are crucial as they provide the lead times. When we aim to optimize supply chains, we need to consider that all decisions are, I would say, lead time dependent. Moreover, these lead times can vary seasonally. For example, there can be busy periods for your suppliers when they’re not as available. These can be seasonal according to the western calendar, but they can also follow the eastern calendar, such as the Chinese New Year, which can add about two weeks to lead times for products coming from China or Asia in general.
Absolutely, and there are also sub-total catches to consider. For instance, if you order a hundred units and your supplier delivers eighty, do you record this information? If you have quality control measures in place where the supplier delivers a hundred but twenty do not pass the quality check, that’s a significant factor. This is especially relevant in the fresh food business, for example, where products like strawberries are easily damaged. You can end up with 20% of the shipments that you receive failing quality control. This can be anticipated, as each year a certain fraction of strawberry shipments will not meet quality control. So, if your need is one hundred, perhaps you should order one hundred and twenty because statistically, you know ahead of time that 20% will not pass quality control.
Kieran Chandler: Right, so the purchase orders and the care taken in noting the dates and quantities where the supplier might make mistakes are all quite crucial.
Joannes Vermorel: Yes, and also it’s this level of granularity that you’re really interested in. This brings me to the last piece of the puzzle: the snapshot of current stock levels. We already know what the purchase history was, and we also know what the sales history was. We have intelligent supply chain scientists here at Lokad. So, do we really need that current snapshot? Can’t they generate that from the other two histories?
Kieran Chandler: In theory, yes, you could reconstruct the stock levels just by knowing exactly how many units came in and how many units went out. That would give you your theoretical stock level at any point in time. The challenge is that inventory inaccuracies happen once in a while. If all you do is retrace the entire history to figure out the stock levels, those inaccuracies would pile up over time to a point where it’s vastly inaccurate.
Joannes Vermorel: That’s correct. In practice, we need inventory snapshots and controls. Firstly, we need to know the current stock levels, because if we don’t know what we have right now, we cannot just rely on the projected future demand to place purchase orders. We have to take into account what we already have. If we already have a mountain of stuff, even if we project very high demand, we might not reorder just because we have enough. So, we need the current stock level, but also the past stock levels. This goes back to the sales history. The sales are not the demand. If you had a stock out, then basically you had no observed demand for probably days, but that’s not because there was no demand, it’s just because you couldn’t serve it. So having the historical stock level is the way to know your historical stock outs, and the situation can be a bit tricky.
People want to buy, let’s say, something like a door handle. They have an apartment and they want to have five door handles that look exactly the same. If they walk into your store and only find three of them that look the same, they are not going to buy three, hoping that they’ll find two identical ones in another store. They’d just prefer to find five that look alike or try another store that can sell them five identical ones. So, stock out is not a binary concept, where it’s either you have exactly zero units and you’re stocked out, or you’re good. Sometimes people are seeking specific quantities, so you have a more nuanced concept of stock out. That’s why you need to keep track of stock levels, and historical stock levels can also be helpful to automatically detect the most likely inventory inaccuracies.
Kieran Chandler: You’ve spoken a lot about the core data sets, the main data sets that Lokad requires. What about those secondary data sets? Is there anything that could also be of interest to us there?
Joannes Vermorel: Yes, the core transactional data set is what we primarily use. The second circle probably involves everything related to promotional activity. Promotions, in the very basic sense of the word, aim to highlight or put forward. It’s not just a matter of pricing. It’s also what you promote on your various channels. For example, an e-commerce company can do promotions that just involve the five products that end up listed permanently in the newsletter sent to the entire customer base. That’s a sort of promotion, even if it doesn’t come with a price discount. It’s a product that appears on the front page of the online store. So, the second circle encompasses everything related to promotions, including advertising. This type of data also probably includes all the e-commerce traffic data they have on the website.
Kieran Chandler: Let’s discuss some of the current clients that Lokad has. Are there any common mistakes that we’ve experienced that future customers could perhaps learn from?
Joannes Vermorel: I think the biggest mistake number one is to choose an ERP system where it’s exceedingly difficult to extract data. That’s a form of vendor lock-in. There are many companies that choose a system where, even if it’s in-house, they have a local server that does everything locally. They think they’re in control because the machine is in their offices.
Kieran Chandler: The reality is that it’s incredibly difficult to extract data because the system hasn’t been properly engineered to make it convenient. For example, extracting anything from QuickBooks in the US market is a massive challenge. I don’t mean any disrespect towards Intuit, they’re a great company, but extracting data is just incredibly difficult.
Joannes Vermorel: Indeed, that’s probably the first mistake. The second mistake is likely not recording gaps of stats, which typically happens with purchase orders. For a smaller company, say with a ten million dollar turnover, your purchase orders could fit on an Excel sheet. However, if you don’t have a better system to record them, the day you want to do any lead time analysis, you’ll struggle because the data is spread across dozens of spreadsheets.
The third mistake might be failing to snapshot stock levels. Nowadays, hard drives are very cheap, yet many companies don’t record historical stock levels. So when we say we need the past stock situation or the past low stock situation, the data has never been recorded. This is unfortunate because we’re talking about a couple of gigabytes of data which should literally cost pennies. There is no reason not to record this data and keep it indefinitely.
Kieran Chandler: To wrap things up, if our listeners were to take away one key point from today’s discussion, what would that be?
Joannes Vermorel: I would say, pay attention to your data. It’s not naturally very expensive, but the best time to start properly recording all the data you need is today. In three years, when you need a few years’ history, if you’ve not started today, you won’t have it. Collecting and maintaining properly organized data is something that you cannot delay, it has to be done now.
As a supply chain manager, you cannot make correct decisions without data. You cannot optimize what you don’t measure, regardless of whether you’re using something as sophisticated as Lokad to optimize your supply chains. This applies not just to the statistical algorithms that Lokad uses, but also to humans. Without data, the decisions you take will not be good. You can’t just guess how many units you’re going to sell for hundreds of thousands of products.
It all starts with choosing the right software to operate your company and making sure you have access to the data. Otherwise, you’ll be locked into the vendor solution that you’ve chosen.
Kieran Chandler: Great. That’s a fundamental insight for someone like me who tends to put things off until tomorrow. That’s everything for this week. We’ll be back again next week with another episode, but until then, we’ll see you soon. Bye for now.
