Rossmann Sales Forecasting
Time series
Machine Learning
Daily sales forecasting for 1,115 stores, 6 weeks ahead: classical statistical models vs. a single global machine learning model.
Problem
Forecast daily sales for 1,115 Rossmann stores 42 days ahead.
Approach
Comparison between classical statistical models fitted per store and a single global machine learning model trained on all stores.
Technologies
Python pandas Time series Machine Learning Walk-forward validation Optuna SHAP Streamlit
Results
A single global model (LightGBM) reached RMSPE 0.1321 and MAPE 9.7% on the test set, beating the specialized SARIMAX in 3 of the 4 series compared. The project also documents where the classical model still wins: the 17 stores with atypical behavior.