Out-of-stock (OOS) events present a real conundrum to consumer goods companies when forecasting demand. On the one hand, you can’t simply ignore previous OOS events if you hope to prevent them in the future, but on the other hand, determining the “true demand” if out-of-stocks had not caused lost sales is a complex undertaking.
Using just aggregate sales, even if it is sell-through vs. sell-in, is insufficient for demand forecasting because historical demand is under-estimated where OOS events have occurred. For example, if a great promo causes an item to go out-of-stock or products are perpetually understocked in certain stores, the actual demand for the product is most likely higher than what the sales numbers reflect.
Those who don’t take that insight into account are doomed to repeat the mistake and under-supply demand as the forecast is translated into shipment and production plans. That’s why out-of-stocks is one of the five critical data types for a best-in-class forecast that we discuss in The No-Nonsense Guide to Forecasting.
So once you have the data, how do you incorporate it into your forecasts? Here, we'll discuss three top methods for treating out-of-stocks in your demand forecast and their pros and cons.
As the name implies, exclusion basically deals with out-of-stocks by removing any data points that might have been affected by OOS from your time series.
Since it completely deletes sales that would have been impacted by OOS, the result is a neat time series that only contains data that is closer to true demand for use in forecasting. The “dirty” data points do not have the opportunity to bias the forecast lower because of uncaptured lost sales, and the methodology is fairly straightforward and easy to implement given the relative infrequency of OOS.
However, OOS events typically do not occur at random. For instance, out-of-stocks may be linked to certain stores, regions, or seasonal patterns where demand is higher than average. These patterns hold critical insights that, if effectively identified, could make a real difference in forecast accuracy, not to mention sales and marketing decisions.
As a result, while exclusion will give you good baseline forecasts because out-of-stocks aren’t very frequent, it is likely to still underpredict the exceptional (critical to profitability) times when OOS tend to occur, such as promos or holidays.
This method is another attempt at creating a clean time series, while avoiding the gaps left by exclusion. Instead, you make a concerted effort to fill those gaps by estimating (i.e., imputing) the true demand levels where there were OOS. It asks the question: "What would sales have been if there had been enough inventory on shelf to fully service demand?"
The main problem a company faces when applying this method is the difficulty in rescaling OOS events to estimate sales if there were not OOS events. To create an accurate estimate of true demand, there has to be enough historical data from similar locations and events that don’t have OOS events in order for you to get a realistic idea of what true demand should be. Otherwise, the estimates are likely to be just guesses with the sheer purpose of filling in the gaps, regardless of the accuracy. This process would introduce bad data that could be misleading, and you wouldn’t even be able to determine whether you are over- or under-predicting true demand.
If you are able to accurately estimate true demand, the advantages of imputation are great. You have a clean, continuous time series that gives you the most flexibility in terms of applying different forecasting models and can also be leveraged for other applications. Setting your team up for success goes back to the data again - the more granular (store and SKU-level) sell-through and out-of-stock data you can get, the more opportunities you have to directly compare similar locations and events and better estimate the imputed values when you have OOS.
3. Categorical Indicator Variable (Flagging)
The third way to incorporate an OOS event into your demand forecast is to actually acknowledge it inside the forecast as a special event. The last two methods took OOS into account by adjusting the sales numbers of the time series; flagging is separate from the sales data and incorporates OOS events with a categorical indicator variable that can go into the model formulation.
Flagging allows you to categorize OOS events, so you can treat different types of OOS differently. For example, an OOS event caused by a retailer consistently under-ordering vs. one caused by a sudden weather event that impacted shipments, like a hurricane, should not necessarily be treated the same way when forecasting demand. It also allows you to differentiate them from true demand at the times where OOS events didn’t occur.
While flagging does more effectively include OOS events, it only does so only at the broad categorical level - not quantitatively like with imputation. Because flags are typically a discrete variable, it is hard to also incorporate the different magnitudes of the impact of each of these types of out-of-stock events, particularly if they are one-off events. This nuance would get lost, so if it can be calculated, a more precise method could yield more accuracy.
Summary of methodologies for incorporating OOS in your forecast
In a follow-up post, we’ll discuss how to choose which methodology to use to incorporate out-of-stocks in your demand forecasts. As you can see, there’s advantages and disadvantages to each one, and your industry, the type of forecasting model you’re using, and data you have access to should all have an impact on your decision. Please sign up down below to stay tuned to our latest updates and news!