Sales Forecasting with XGBoost and Stacking

This project explores different modeling strategies for forecasting sales, using XGBoost regressors and a stacking ensemble approach.

Project Structure

• Model 1: One global model trained on all historical data.
• Model 2: One global model trained only on recent data.
• Model 3: Separate model for each product family, trained on all historical data.
• Model 4: Separate model for each product family, trained only on recent data.

Key Results

Model	RMSLE	Notes
Model 1	0.6333	Single global model on full historical dataset
Model 2	0.5277	Single global model on recent data split
Model 3	0.4128	Separate model per family on full historical data
Model 4	0.4152	Separate model per family on recent data split

Approach

• Feature engineering included promotions, holidays, and oil price data.
• Models were trained using XGBoost regressors.
• Stacking was applied by using family-level model outputs as meta-features for a second-stage model.
• Evaluation metric: Root Mean Squared Log Error (RMSLE).