Stockout-corrected baseline forecast
Exponential smoothing per product and store, or per group for slow sellers, fitted to a sales history in which stockout periods have been replaced with an estimate of the demand they hid.
Part of the recommended approachTechnique: Statistical Time Series Forecasting
This is the forecast that feeds the reorder points for most of the range.
Before fitting anything, it marks each period in which a product was out of stock in a store and replaces that period’s sales with an estimate of demand, for example the average of the surrounding periods, keeping the recorded figure whenever it is higher.
It then fits exponential smoothing to the corrected history, per product and store where sales are dense enough and per group where they are not, testing a Croston-type variant against simple exponential smoothing for the slowest sellers.
Forecasting: Principles and Practice describes exponential smoothing as producing reliable forecasts quickly for a wide range of series, which is why it is the fast default to build first and the benchmark any more complex forecast has to beat.
It is the wrong choice as the only forecast for promotional or event-driven products, because their demand is driven by things that are not in their own history.
It also cannot see a phantom stockout, since the inventory record says stock was present, so its correction misses exactly the periods phantom inventory screening is built to find.
Strengths for this problem
- Breaks the loop in which every stockout lowers the next forecast and makes the next stockout more likely.
- Fits every product and store combination in minutes on ordinary hardware, so the whole range can be refreshed every cycle.
- Each forecast decomposes into level, trend, and weekly or yearly seasonality that a planner can inspect and argue with.
- Croston-type variants, the most common method for intermittent demand, are available for the many slow sellers whose weekly sales are mostly zeros.
Limitations for this problem
- Does not anticipate promotions, price changes, or events unless their effect is added separately.
- A single store's history for a slow seller is too thin to forecast alone, so it needs to be grouped by product hierarchy or store cluster and spread back down.
- The stockout correction is only as good as the stockout flag, which is itself wrong wherever the inventory record is wrong.
Data needed
- Daily or weekly units sold per product and store.
- Daily on-hand quantity and recorded out-of-stock events, to flag the periods whose sales are censored.
- Product hierarchy and store clusters, to group slow sellers.
- Holiday and store closure calendar.
How it is measured
- Scaled forecast error, such as mean absolute scaled error, on stockout-corrected demand rather than raw sales.
- Forecast bias per category, tracked separately from error, since a consistent under-forecast is what causes repeat stockouts.
- Improvement over a seasonal naive forecast, which any baseline must beat to be worth running.
Pitfalls
- Fitting to raw sales without the correction, so the forecast inherits the stockouts it was meant to prevent.
- Adjusting stockout periods downwards when the estimate is lower than recorded sales, which removes real demand.
- Reporting MAPE at product and store level, where zero-sales weeks make it undefined or dominated by the slowest sellers.
- Running one method for every series without testing a Croston-type variant against simple exponential smoothing on the slow sellers.
Human oversight
The planner reviews forecasts for promotions, launches, and products with a recent range or price change, and can override them with a logged reason. Analysts review the stockout correction rule when it changes, since it moves every forecast in the range.