It won on the steady sellers and lost on the rest
On one Walmart store, a machine-learning forecast was 3.9 percent worse than a 28-day moving average on absolute error, and the worse forecast for 58 percent of the 823 items. Split by demand class, the result is two opposite ones: it wins on the steady sellers and loses by about 10 percent on the intermittent and lumpy items, which are most of the store. The usual way to choose a forecast is to…
A Walmart store's machine-learning forecast proved 3.9% worse than a 28-day moving average on average error, incorrectly performing better on 58% of the 823 items. This disparity emerged due to the model's superior performance on steady sellers and poor performance on intermittent and lumpy items, which make up most of the store's inventory.
The usual method for selecting a forecast based solely on the smallest error number can lead to incorrect choices for most of the catalogue. Utilizing forelint, an open-source library for auditing forecast backtests, this analysis demonstrated that the moving average was the better choice for 77% of the store's items, particularly the intermittent and lumpy ones.
This insight highlights the importance of considering demand class when selecting forecasts, rather than relying on a single model for all items.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.