Advanced
Open
Pro
Forecasting a SKU With No Sales History
Part of the ML System Design Interview path →
Part of the Advanced Data Science & Experimentation path →
The retailer launches about 12 new SKUs per half-year. A global gradient boosting model that handles young series through missing lags, an age feature, and a scale taken from category-by-store peers gives, on the cells of young series and series launching inside the horizon (240 series, so indicative only):
| Method | WAPE | Bias |
|---|---|---|
| Global model (peer scale, age features) | 0.764 | -23.3% |
| Flat peer level (category by store mean) | 0.886 | -23.8% |
| Own mean since launch | 0.839 | -65.1% |
| Seasonal naive | 0.955 | -79.4% |
For forecasts made before launch, with no history at all, the global model's WAPE is 1.217 and the peer level's is 1.370. The true-mean floor on these cells is 0.580.
- What does a WAPE above 1 mean here, and what should the first order for a new item be based on?
- Why is the own-mean baseline so biased?
- Propose a better statistical treatment for the first weeks.
- How do you avoid fooling yourself in the evaluation, given only 12 launches?
Share this question