CERTAIN INVESTIGATION AND PREDICTION OF BIGMART SALES USING MACHINE LEARNING TECHNIQUES

Abstract

Accurate sales forecasting is critical for retail optimization, yet many predictive models struggle with the inherent noise and categorical complexity of retail datasets. This study presents a rigorous investigation into machine learning techniques for predicting sales using the Big Mart dataset. Unlike previous narrative studies, we implement a reproducible pipeline incorporating median-imputation for missing values, one-hot encoding for categorical features, and a 5-fold cross-validation protocol to ensure model stability. We evaluate four distinct architectures: Linear Regression, Polynomial Regression, Decision Trees, and XGBoost. Our results demonstrate that XGBoost significantly outperforms traditional models, achieving a Root Mean Squared Error (RMSE) of 0.0231 and a Mean Absolute Error (MAE) of 0.018 after systematic hyperparameter tuning. Beyond predictive accuracy, we provide a feature importance analysis revealing that Item MRP and Outlet Type are the primary drivers of sales variance. This work contributes a validated framework for retail forecasting that balances predictive power with business interpretability, offering actionable insights for inventory management and strategic decision-making.