Build one model that works across all four shops' data. Can what it learns on one help on the others?
Task
Forecast weekly units sold for every article in each shop, 13 weeks ahead.
Data
Sales history from four shops: Set A, Set B, Set C and Set D. They share no stores or products. Get the data.
Test
We hide each shop's last 13 weeks. You forecast them from the weeks before.
Final test
We score your model on a shop's data that is not in this set. So test it on a shop it never trained on.
Score
(1 − MASE) × 100. Higher is better; 0 means you miss by one typical week-to-week move. AutoARIMA averages 32.0. Beat it.
Question
Does what you learn on one shop help on another? Show when it helps and when it hurts.
Extra credit
Show that the other sets helped. Run your model twice: own data only, then with the other sets. Show both scores. If borrowing helped, say why. If it hurt, say when. We weigh this when we pick the winner.
Does borrowing another shop's shape help?
Each shop borrows Set C's yearly shape, built only from weeks before its own holdout.
Borrowed Set C season: score -15.8 (own-only -52.5, naive -28.6)
Borrowed Set C season: score -129.8 (own-only 6.9, naive -2.3)
Actual Own-only top-down (MSTL) Borrowed Set C season Naive (same as last week)
It helped one store and hurt another. When does it help?