Retail Sales Forecasting — Case Study

An end-to-end time series forecasting pipeline predicting weekly sales at the store/department level, built for a retail demand-planning use case. The project covers data understanding, EDA, feature engineering, statistical and machine learning forecasting models, model evaluation, and business recommendations.

Business Problem

A retail chain wants to improve demand planning and inventory allocation by forecasting weekly sales across 45 stores and 81 departments (3,323 unique store-department series), accounting for seasonality, holiday spikes, markdown-driven fluctuations, and store-level performance differences.

Repository Structure

.
├── TimeSeriesAnalysis_2.ipynb      # Full analysis notebook (EDA → modeling → forecasts)
├── clean_data.zip                  # Source dataset — unzip before running
├── forecast_output.csv             # Final predictions (Date, Store, Dept, Actual, Predicted)
├── REPORT.pdf                      # Executive summary, insights & recommendations
├── requirements.txt
└── README.md

Dataset

Weekly sales data for 45 stores / 81 departments (2010–2012), including holiday flags, store type/size, temperature, fuel price, CPI, unemployment, and promotional markdown information. Target variable: Weekly_Sales.

Methodology

1. Data Understanding — schema check, null/duplicate audit, granularity analysis (panel data: 3,323 distinct store-dept time series).
2. EDA — trend/seasonality decomposition, holiday impact quantification, store/department comparison, correlation analysis.
3. Feature Engineering — calendar features (week/month/quarter), lag features (1, 2, 4, 13, 26, 52 weeks), rolling statistics, a corrected "true peak week" holiday flag (see Key Findings).
4. Feature Selection — correlation/multicollinearity check; full engineered feature set retained (no problematic collinearity found; tree models are robust to correlated predictors).
5. Modeling — two model families, five total models:
   • Statistical: ARIMA(2,0,2), SARIMA(0,0,1)(0,1,1,52)
   • Machine Learning: Linear Regression, Decision Tree, Random Forest, XGBoost (hyperparameter-tuned via RandomizedSearchCV with TimeSeriesSplit)
6. Evaluation — MAE, RMSE, %RMSE, MAPE, and WMAE (holiday-weighted, matching the original Walmart competition standard); all splits are strictly chronological to prevent leakage.
7. Forecast Generation — final predictions exported to forecast_output.csv.

Key Findings

• Strong annual seasonality (Thanksgiving/Christmas-driven); the dataset's official "Christmas" flag is mistimed — the true peak occurs the week before (week 51), at roughly 2x the flagged week's sales.
• lag_52 (same week, one year ago) is the single strongest predictive feature across all tree-based models.
• Store Size/Type are meaningful sales drivers; macroeconomic indicators (CPI, fuel price, unemployment) show weak linear correlation with sales.
• Markdown impact on sales is confounded with an overall time trend and should be treated cautiously.

Model Results

Model	MAE	MAE (% of mean sales)	RMSE	%RMSE	MAPE	R²
Random Forest (selected)	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 (company-level baseline)	—	—	—	3.36%	2.67%	—

Random Forest was selected as the final model — lowest error across every metric, with strong generalization from ensembled deep trees on a dataset dominated by one highly informative seasonal-lag feature.

Tech Stack

Python · pandas · NumPy · scikit-learn · XGBoost · statsmodels · matplotlib · seaborn

Setup & Usage

pip install pandas numpy scikit-learn xgboost statsmodels matplotlib seaborn
jupyter notebook TimeSeriesAnalysis.ipynb

Run all cells top to bottom. Outputs (plots, metrics, forecast_output.csv) are generated in place.

Full Report

See Retail_Sales_Forecasting_Report_.pdf for the executive summary, detailed insights, and business recommendations.