🛍 Rossmann Retail Sales Forecasting: Causal Impact Analysis and Explainable Machine Learning

---

📖 Project Overview

In this project, we develop a complete data science pipeline to forecast daily sales for a major retail chain (Rossmann), quantify the true business impact of promotional campaigns, and provide full explainability of model decisions using modern AI techniques like SHAP.

We combine traditional time series forecasting, machine learning regression models, manual causal impact analysis, and SHAP explainability into a professional, real-world forecasting solution.

---

❓ Problem Statement

Retailers like Rossmann face highly volatile daily sales influenced by promotions, competition, holidays, and store characteristics.
Traditional forecasting models (like ARIMA or Prophet) can predict trends but cannot quantify the business impact of promotions or explain why predictions are made.

Our goals:

• Predict future sales accurately.
• Quantify how much promotions actually lift sales.
• Make machine learning model predictions interpretable to business users.

---

📚 Dataset Summary

Rossmann Sales Dataset from Kaggle:

• ~1 million rows of daily sales data (2013-2015).
• Includes store metadata (type, assortment, competition, promotions).
• Data for 1115 stores across different regions.

---

🏗 Project Architecture

rossmann_forecasting/
├── data/                      # Raw and cleaned datasets
├── notebooks/                  # Exploratory notebooks per step
├── src/                        # Modular Python scripts
├── outputs/                    # Saved plots, causal reports
├── requirements.txt            # Python dependencies
├── README.md                   # Project documentation
├── LICENSE                     # MIT License

---

🛠️ Approach and Methods

1. Data Preprocessing

• Merged sales and store metadata.
• Parsed dates into year, month, day, and week features.
• Removed closed stores and zero-sales records.

2. Feature Engineering

• Lag features (7, 14, 30 days).
• Rolling mean features (7-day, 30-day).
• Promo duration features (cumulative running promo days).
• Competition open time features (months since competitor store opened).

3. Baseline Statistical Modeling

• Prophet and SARIMA time series models built for each store.
• RMSE performance recorded.

4. Machine Learning Forecasting

• Random Forest and XGBoost regression models.
• Feature set: lagged sales, promotions, competition, seasonality.
• Evaluation metrics: RMSE, MAE.

5. Causal Impact Analysis (Manual)

• Built counterfactual sales model using linear regression on pre-promo period.
• Measured uplift during promotions by comparing predicted vs actual sales.
• Quantified average and relative lift in sales.

6. Model Explainability

• SHAP values for model interpretation (feature importance and individual impact).
• Partial Dependence Plots (PDP) for top features.

---

📊 Key Results

Model	RMSE (Test Data)
Prophet (Baseline)	~780
SARIMAX	~1285
Random Forest Regressor	357.14
XGBoost Regressor	368.47

✅ Machine learning models outperformed traditional baselines, reducing RMSE by ~50%.

---

🎯 Causal Impact Findings

• Average Daily Sales Lift from Promotions: $39.96
• Relative Lift: +0.90%
• Promotions showed a positive but moderate impact on sales volumes.

✅ Sales behavior during promotions was significantly higher than model-predicted counterfactual sales without promotions.

---

🖼️ Visualizations (Located in /outputs/plots/)

• SHAP Summary Plot (shap_summary.png): Global feature importance across sales forecasts.
• Random Forest Feature Importance Plot (rf_feature_importance.png): Top predictive features for sales.
• Partial Dependence Plots (pdp_plots_rf.png): Impact of Promo, Day of Month, and Sales history.
• Causal Impact Plot (causalimpact_plot.png): Actual vs Predicted Sales During Promotion.

---

⚙️ Technologies Used

Tool	Purpose
Python 3	Main programming language
pandas, numpy	Data manipulation
scikit-learn	Machine learning models
XGBoost	Gradient boosting
Prophet	Time series forecasting
statsmodels	ARIMA modeling
SHAP	Model explainability
Matplotlib, Seaborn	Data visualization

---

🚀 How to Run This Project

1. Clone the repository:

   git clone https://github.com/yourusername/rossmann_forecasting.git
   cd rossmann_forecasting

2. Install dependencies:

   pip install -r requirements.txt

3. Run notebooks sequentially:

   • 01_EDA_FeatureEngineering.ipynb
   • 02_StatisticalModels_ARIMA_Prophet.ipynb
   • 03_MLModels_XGBoost_RF.ipynb
   • 04_CausalImpact_Analysis.ipynb
   • 05_Explainability_SHAP_PDP.ipynb

✅ All modular scripts (src/) are available if you want to pipeline this project.

---

🔮 Future Work

• Automate hyperparameter optimization (Optuna).
• Build live forecasting dashboard using Streamlit.
• Expand causal analysis to include external events (e.g., holidays, weather).
• Model multiple stores using hierarchical time series (HTS).

---

💜 License

This project is licensed under the MIT License — free to use and modify.

---