Rossmann Sales Forecasting

The setup

Rossmann's store managers forecast daily sales up to six weeks ahead. The model gets exactly what they would have: store metadata, the calendar, promotions, school and state holidays, and competition timing — nothing measured on the day being forecast. The holdout is the last six weeks of the data (40,282 open store-days), models train on everything before it, and scoring uses RMSPE, the competition's own metric.

What the numbers say

Model	RMSPE ↓	R²
Naive: same weekday, 364 days earlier	0.1718	0.8306
Naive: store × weekday × promo median	0.1449	0.8585
Ridge	0.1567	0.8551
LightGBM	0.1235	0.9195

A good baseline is hard to beat. A lookup table of store × weekday × promo beats the linear model outright; most of the signal is which store, which day, promo or not. LightGBM is the only model that clears it — a 14.8% reduction in RMSPE — and 81% of its split gain comes from two store-level encodings learned from the training period alone.

Keeping the forecast honest

The first version of this project scored R² 0.965 with same-day customer counts among its inputs — a number no forecast has six weeks ahead. The rebuild removes it and measures its effect instead: added back, it roughly halves the error (RMSPE 0.0628), which is exactly why it cannot be used. A test now fails the pipeline if any feature for a given day reads data from that day or later.

What I am building next

Per-store error analysis on the stores the model gets most wrong, and a rolling-origin evaluation across several six-week windows instead of one.