Rossmann demand forecast

Daily sales for 1,115 drugstores, up to a 42 day horizon. Gradient boosted trees on log1p(sales), 41 features, 407 boosting rounds. Scored once on a test window that no training step used.

Model MAPE

8.63%

test window

Baseline MAPE

18.56%

seasonal median

Error reduction

53.5%

against baseline

RMSE

881

euros per day

Try a forecast

Store

From

Example

forecast actual promotion day

Model card

Field	Value
Algorithm	LightGBM, 407 rounds, 41 features
Target	log1p(sales), trading days only
Trained through	2015-06-13
Servable range	2015-06-14 to 2015-09-17
Stores	1,115
Test RMSE / MAPE / WAPE	881 / 8.63% / 8.37%
Baseline RMSE / MAPE / WAPE	1,671 / 18.56% / 17.29%

The upper date bound is 48 days after the last day of sales history. Every feature is at least 48 days old by design, so the service needs no recursive feedback. It also cannot see past that date, and returns a 400 if you ask it to.

Endpoints

/docs — OpenAPI interface
/health — model provenance and metrics
/stores — store IDs this service can forecast
GET|POST /predict — store_id, start_date, end_date