Forecast design
42-day direct horizon
Training ended on 19 June 2015. The period from 20 June to 31 July 2015 was retained as the unseen test period.
Overall performance
Model comparison
Lower values indicate better forecasting accuracy.
Store analysis
Actual vs forecast
Consistency across stores
Store-level MAPE
The best value for each store is highlighted.
| Store | Seasonal Naive | Holt-Winters | HistGradientBoosting | Best Model |
|---|
Preferred model
HistGradientBoosting
It achieved the lowest MAE, RMSE and MAPE overall, and the lowest store-level MAPE for all 10 selected stores.
Important limitation
Controlled evaluation
Results are based on 10 selected stores and one holdout window. Wider deployment should be validated across more stores and periods.