M5 Forecasting - Walmart Sales

Objective

Predict the unit sales of 30,490 Walmart products for the next 28 days.

Data Source

Kaggle M5 Forecasting Accuracy

Project Structure

PREDICTIVE-ANALYTICS-PROJECT/
├── configs/
│   └── config.yaml        ← Edit this to change features, model params, paths
├── data/
│   ├── calendar.csv
│   ├── sales_train_validation.csv
│   ├── sales_train_evaluation.csv
│   ├── sell_prices.csv
│   └── sample_submission.csv
├── notebooks/
│   └── feature_importance_analysis_v1.ipynb  ← Run once to decide feature list
├── src/
│   ├── preprocessing.py   ← Load CSVs, melt, merge, memory optimisation
│   ├── features.py        ← All feature engineering logic
│   ├── models.py          ← Model definitions, training, ensemble
│   └── train.py           ← Main entry point (run this)
├── .gitignore
├── README.md
└── requirements.txt

How to Run

python src/train.py
# or specify a different config:
python src/train.py --config configs/config.yaml

How to Change Things

What you want to do	Where to change
Change which features to use	configs/config.yaml → features.use
Change lag / rolling windows	configs/config.yaml → features.lags / features.rolling_means
Change model hyperparameters	configs/config.yaml → model.lgbm or model.xgb
Switch to XGBoost	configs/config.yaml → model.active_model: "xgb"
Use ensemble (LGBM + XGB)	configs/config.yaml → model.active_model: "ensemble"
Add a new model type	src/models.py → add get_<name>_model() and register in get_model()
Add a new feature	src/features.py → add to build_features(), then add name to configs/config.yaml
Change data or output path	configs/config.yaml → data.path / output.submission_path

Team Workflow

Step 1 — Feature selection (run once):

• Open notebooks/feature_importance_analysis_v1.ipynb
• Run it to see which features matter
• Update configs/config.yaml → features.use accordingly Step 2 — Experiment with models:
• Change model.active_model in configs/config.yaml
• Run python src/train.py
• Compare WRMSSE scores Step 3 — Try ensemble:
• Set model.active_model: "ensemble" in config
• Adjust model.ensemble.weights to weight better-performing models higher