← Back to projectsMachine learning / Forecasting
Rossmann sales forecasting
A LightGBM pipeline for the Rossmann Kaggle competition, with log-transformed targets, Optuna tuning, and shared feature logic across notebooks.
Measured signal0.1199held-out RMSPE
Measured signal49/56/63day windows
Measured signal3pipeline stages
01 / The problem
Retail sales are shaped by promotions, holidays, store characteristics, and temporal patterns. A notebook can produce a strong score while still making it easy to leak future information or mismatch predictions to rows.
02 / The approach
The project separates feature engineering, modeling, and inference, shares feature logic across stages, uses chronological validation, and keeps identifiers in the pipeline to avoid silent row mismatches.
03 / What shipped
- Lag and rolling features with calendar, promotion, and competition signals
- Log-transformed target with Optuna hyperparameter search
- One shared feature implementation across train, validation, and inference
- A reproducible notebook handoff through saved artifacts
04 / Outcome
It is a compact example of disciplined tabular ML: not the loudest project, but a useful baseline for showing how reliable experiments are assembled.
A clean experiment is one where the next person can tell exactly what happened.