BEYOND GRADIENT BOOSTING: A COMPARATIVE STUDY OF XGBOOST AND FEEDFORWARD NEURAL NETWORKS FOR RETAIL SALES CLASSIFICATION USING THE BIG MART DATASET

Abstract

Accurate sales forecasting sits at the center of retail competitiveness, shaping decisions on inventory positioning, staffing, pricing, and promotional timing. This study investigates whether a modern gradient boosting algorithm, XGBoost, or a deep feedforward neural network (FNN) is better suited to classifying retail sales performance using the widely studied Big Mart dataset [1]. Following a structured preprocessing pipeline that addressed missing values in Item_Weight and Outlet_Size, encoded categorical attributes, and standardised numerical features, the continuous sales variable was converted into a binary classification target distinguishing high performing from low performing items. Both models were trained on an 80:20 split and evaluated using accuracy, precision, recall, confusion matrices, and ROC-AUC scores. On the held-out validation set, XGBoost reached an accuracy of 89.6% and an AUC of 0.926, while the FNN, trained with dropout regularisation and the Adam optimiser, achieved 91.2% accuracy and an AUC of 0.939. Feature importance analysis identified Item_Visibility, Outlet_Age, and Item_Weight as the strongest predictors of sales class. The results indicate that although XGBoost remains faster to train and easier to interpret through its native feature importance rankings, the FNN captured non-linear interactions among product and outlet attributes more effectively, yielding a modest but consistent performance advantage. The paper closes with a discussion of the limitations tied to dataset size and the absence of temporal features, and it proposes hybrid modelling, explainability tooling, and real-time deployment as promising directions for future retail analytics research.