An end-to-end retail demand forecasting project that predicts weekly sales across 45 stores, 81 departments, and 420K+ records using statistical time-series models and machine learning. The project combines business analytics, feature engineering, predictive modelling, and model evaluation to support inventory planning and demand forecasting.
- 📊 420K+ weekly sales records
- 🏬 45 Stores | 81 Departments
- 📅 3 Years of historical sales data
- 🤖 Compared 6 forecasting models
- 🏆 Best Model: Random Forest (R² = 0.987)
- 📈 End-to-end Time Series & Machine Learning pipeline
- 📦 Business use case: Retail Demand Forecasting & Inventory Planning
Retail businesses must accurately forecast product demand to optimize inventory, workforce planning, and supply chain operations. This project predicts weekly sales across multiple stores and departments while accounting for seasonality, holiday effects, promotional markdowns, and store-specific characteristics to improve business decision-making.
.
├── TimeSeriesAnalysis_2.ipynb # Complete analysis notebook
├── clean_data.zip # Dataset (extract before running)
├── forecast_output.csv # Final predictions
├── Retail_Sales_Forecasting_Report.pdf
├── requirements.txt
└── README.md
The project uses the Walmart Retail Sales dataset containing weekly sales information from 2010–2012.
| Attribute | Description |
|---|---|
| Stores | 45 |
| Departments | 81 |
| Store–Department Series | 3,323 |
| Total Records | 420K+ |
| Time Period | 2010–2012 |
| Target Variable | Weekly_Sales |
Additional variables include:
- Holiday Flag
- Store Type
- Store Size
- Temperature
- Fuel Price
- CPI
- Unemployment
- Markdown Features (MarkDown1–5)
The forecasting pipeline consists of the following stages:
- Dataset exploration
- Missing value analysis
- Duplicate detection
- Panel data identification
- Data quality validation
- Weekly sales trend analysis
- Seasonal decomposition
- Holiday impact analysis
- Store and department comparison
- Correlation analysis
- Distribution analysis
Engineered predictive features including:
- Calendar Features (Week, Month, Quarter)
- Lag Features (1, 2, 4, 13, 26, 52 weeks)
- Rolling Mean & Rolling Standard Deviation
- Holiday Features
- Store Characteristics
- Economic Indicators
- Correlation analysis
- Multicollinearity assessment
- Feature importance analysis
The complete engineered feature set was retained since no significant multicollinearity affecting tree-based models was observed.
- ARIMA (2,0,2)
- SARIMA (0,0,1)(0,1,1,52)
- Linear Regression
- Decision Tree
- Random Forest
- XGBoost (RandomizedSearchCV + TimeSeriesSplit)
Models were evaluated using chronological train-test splits to avoid data leakage.
Evaluation metrics:
- MAE
- RMSE
- MAPE
- WMAE
- R² Score
The final model was used to generate weekly sales forecasts for every Store–Department combination.
| Model | MAE | MAE (% of Mean Sales) | RMSE | %RMSE | MAPE | R² |
|---|---|---|---|---|---|---|
| 🏆 Random Forest | 1,226.53 | 7.70% | 2,496.83 | 15.68% | 208.64% | 0.987 |
| XGBoost (Tuned) | 1,339.64 | 8.41% | 2,810.39 | 17.65% | 185.48% | 0.984 |
| Decision Tree | 1,489.07 | 9.35% | 3,179.38 | 19.97% | 78.33% | 0.979 |
| Linear Regression | 1,613.06 | 10.13% | 3,007.43 | 18.89% | 864.57% | 0.981 |
| SARIMA (Statistical Baseline) | — | — | — | 3.36% | 2.67% | — |
Evaluation Metrics
- MAE: Mean Absolute Error
- RMSE: Root Mean Squared Error
- %RMSE: RMSE expressed as a percentage of mean sales
- MAPE: Mean Absolute Percentage Error
- R²: Coefficient of Determination
Random Forest achieved the lowest forecasting error and the highest predictive performance among all evaluated machine learning models, making it the final model selected for demand forecasting. SARIMA served as the statistical forecasting baseline for comparison.
- Holiday periods generate the highest sales demand, particularly during Thanksgiving and the week preceding Christmas.
- lag_52 (sales from the same week in the previous year) is the most influential forecasting feature.
- Store Size and Store Type contribute significantly more to forecasting accuracy than macroeconomic indicators.
- Promotional markdown effects are partially confounded with long-term sales trends and should be interpreted cautiously.
- Random Forest consistently outperformed all other statistical and machine learning models.
The developed forecasting pipeline can help retailers:
- Improve demand planning
- Optimize inventory allocation
- Reduce stock-outs and excess inventory
- Support staffing and workforce planning
- Improve promotional planning
- Enable data-driven operational decision-making
| Category | Technologies |
|---|---|
| Programming | Python |
| Data Analysis | Pandas, NumPy |
| Machine Learning | Scikit-learn, XGBoost |
| Time Series | Statsmodels |
| Visualization | Matplotlib, Seaborn |
| Development | Jupyter Notebook |
The project produces:
- Weekly sales forecasts
- Forecast performance metrics
- Feature importance analysis
- Business insights
- Forecast output file (
forecast_output.csv) - Executive report with recommendations
Clone the repository:
git clone https://github.com/your-username/Retail-Sales-Forecasting.git
cd Retail-Sales-ForecastingInstall dependencies:
pip install -r requirements.txtExtract the dataset:
clean_data.zip
Run the notebook:
jupyter notebook TimeSeriesAnalysis_2.ipynb- Deploy the forecasting pipeline using Streamlit
- Build an interactive Power BI dashboard
- Compare with Prophet and LightGBM
- Explore deep learning models (LSTM)
- Automate periodic model retraining
The complete project report containing methodology, visualizations, model evaluation, business insights, and recommendations is available in:
📄 Retail_Sales_Forecasting_Report.pdf
Siddharth Gupta
B.Tech | Netaji Subhas University of Technology (NSUT)
⭐ If you found this project useful, consider giving the repository a star.