End-to-end time series forecasting on Rossmann Store Sales — 1M rows, 1115 German drugstores, 3 years of daily data (2013-2015). Three models compared with walk-forward cross-validation, plus anomaly detection on residuals.
| Model | Test MAPE | Test RMSE | CV MAPE |
|---|---|---|---|
| LightGBM | 8.20% | 453 | 9.09% |
| Prophet | 17.16% | 879 | 14.04% |
| SARIMA | 17.95% | 900 | N/A |
Winner: LightGBM — 2x better accuracy than classical statistical models.
- LightGBM dominates — MAPE 8.20% vs Prophet 17.16% and SARIMA 17.95%
- Top predictive features:
Sales_lag_1(yesterday's sales),Week,Promo - Promo impact: sales increase ~37% on average during promotional periods
- December seasonality: sales peak 60-80% above baseline (Christmas effect)
- Anomaly detection: 14 anomalies flagged (1.8% of test set), 8 of them on Mondays
- Anomaly correlation: high-value anomalies correlate with overlapping promo + holiday windows
Demand forecasting drives inventory, staffing, and revenue planning decisions in retail and supply chain. I wanted to compare a classical statistical approach (SARIMA), a modern automated approach (Prophet), and a modern ML approach (LightGBM) on the same dataset with the same evaluation methodology — to see where each shines and which one I would actually deploy.
The honest answer: for daily retail data with promotional cycles and lag-based autocorrelation, modern gradient boosting beats classical time series methods substantially, mostly because it can use lag features and external regressors (promo, holidays) jointly.
Random K-fold leaks future into past on time series data — a model trained on future days will trivially predict the past. We use rolling-origin walk-forward CV instead:
Fold 1: train 2013-01 → 2013-12 | test 2014-01
Fold 2: train 2013-01 → 2014-01 | test 2014-02
Fold 3: train 2013-01 → 2014-02 | test 2014-03
...
Reported CV MAPE is averaged across all folds. SARIMA didn't run cross-validation due to fit time (single model takes ~10 min per fold per store).
- Lag features:
Sales_lag_1,Sales_lag_7,Sales_lag_30 - Rolling statistics: 7-day and 30-day moving averages and standard deviations
- Calendar: day-of-week, week-of-year, month, quarter, year
- Promo: binary flag + days-since-promo started
- Holidays: state holiday flag, school holiday flag, days-to-next-holiday
- Store features: store type, assortment, competition distance, promo2 participation
Residual-based approach: for each model, compute residual = actual - predicted. Days where |residual| > 3 × σ(residuals) flagged as anomalies. Cross-referenced against promo and holiday windows to separate "real" anomalies (data issues, stockouts) from explainable outliers (mega-promo days).
Rossmann Store Sales (Kaggle) — 1 million rows, 1,115 stores, 3 years of daily sales.
Data not included in this repo (~30MB). Download from Kaggle and place CSVs in
data/.
- Python 3.11+
- LightGBM — gradient boosting (winner)
- Prophet — Facebook's automated forecasting
- statsmodels — SARIMA classical statistical approach
- Plotly — interactive visualizations
- pandas, numpy, scikit-learn — data wrangling and metrics
demand-forecasting/
├── notebooks/
│ ├── 01_eda.ipynb # Exploratory Data Analysis
│ ├── 02_feature_engineering.ipynb # Lags, rolling stats, calendar features
│ ├── 03_prophet_model.ipynb # Prophet + walk-forward CV
│ ├── 04_arima_model.ipynb # SARIMA fitting
│ ├── 05_lightgbm_model.ipynb # LightGBM + feature importance
│ └── 06_comparison_and_anomaly.ipynb # Model comparison + anomaly detection
├── data/ # Not in repo - download from Kaggle
│ ├── train.csv
│ ├── test.csv
│ ├── store.csv
│ └── sample_submission.csv
├── docs/
│ ├── model_comparison.png
│ ├── feature_importance.png
│ └── anomalies.png
├── requirements.txt
├── .gitignore
└── README.md
git clone https://github.com/nurbolsultanov/demand-forecasting.git
cd demand-forecastingpython -m venv venv
# Windows
venv\Scripts\activate
# Mac/Linux
source venv/bin/activatepip install -r requirements.txtIf prophet installation fails on Windows, try:
pip install prophet --no-build-isolationor use conda: conda install -c conda-forge prophet.
- Go to kaggle.com/c/rossmann-store-sales/data
- Click Download All (requires Kaggle account)
- Unzip into
data/folder
jupyter notebookOpen notebooks 01_eda.ipynb through 06_comparison_and_anomaly.ipynb and run cells sequentially.
Lag features and promo flags dominate over pure seasonality components — confirming that daily retail data is strongly autocorrelated and promo-driven, not just seasonal.
Anomalies cluster around overlapping promo + holiday windows. Most flagged days are explainable; a handful are likely data issues worth investigating.
- Single-store focus in detailed analysis. Comparison done on a representative store; full multi-store deployment would require per-store hyperparameter tuning or hierarchical modeling.
- No external regressors. Weather, competitor pricing, and local events not included — they would likely lift LightGBM further.
- SARIMA didn't run CV. Walk-forward CV across folds was prohibitively slow for SARIMA on this dataset. Reported test MAPE is on a single holdout window.
- Anomaly detection is residual-based. A more rigorous approach would use isolation forests or seasonal-trend decomposition. Current method works well for residual outliers but won't catch sustained regime shifts.
- Hierarchical forecasting across all 1,115 stores
- Hyperparameter tuning via Optuna
- Ensemble of LightGBM + Prophet for robustness
- Streamlit dashboard wrapping forecasts and anomalies for interactive exploration
- Probabilistic forecasts (quantile regression) instead of point estimates
- External regressors: weather data, competitor pricing
Built by Nurbol Sultanov — Data Analyst in Los Angeles working in credit risk, fraud detection, and forecasting.
MIT — feel free to fork, modify, and use for your own forecasting projects.