Planning · 08 / 20
Statistical forecast engine
Four forecast models, switchable per SKU, with accuracy tracked alongside.
What actually changes
Stop applying one growth assumption to SKUs that behave nothing alike.
A single flat growth rate gets applied across the board, even though some SKUs are seasonal and others are flat or erratic.
See a tailored forecast and confidence band for every individual SKU
Switch between forecast models per SKU and see which one actually fits its demand shape
Track forecast accuracy over time instead of finding out you were wrong after the stockout
What was built
A per-SKU forecast with confidence bands, four switchable models, and live accuracy stats.
Result
4
models, one switch
The mechanics
How this actually runs, step by step.
- 1104 weeks of history are decomposed into trend, seasonality, and residual.
- 2Four models generate a forecast plus 80%/95% confidence bands.
- 3Accuracy stats (MAPE, WMAPE, bias, tracking signal) are computed per model, side by side.
- 4Switching the model swaps the forecast line and its accuracy panel together.
Holt-WintersCrostonMAPE / WMAPE