LAUSR

ABSTRACT Sales forecasting is a critical component of supply chain management and retailer-manufacturer operations, enabling manufacturers to anticipate future demand for optimized production planning while helping retailers manage inventory and capital expenditures efficiently. Forecasting methods range from statistical models and human-driven planning to hybrid approaches, depending on organizational needs. This study proposes a TabNet-based deep learning model for predicting converted sales in the Big Bazaar retail chain. To enhance prediction accuracy, products are categorized into three distinct groups: (i) perishable and fresh foods, (ii) packaged and processed foods, and (iii) household and miscellaneous items using three clustering techniques: K-Means, Hierarchical Clustering (HC), and Partitioning Around Medoids (PAM). The dataset consists of 14,204 data points, split into 70% training and 30% testing for model evaluation. The study employs Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) as performance metrics. The predictive efficiency of sales forecasting is assessed by comparing multiple models like two ANNs with different hidden layer configurations, two regression models, three DNNs architectures, and TabNet. Experimental results demonstrate that both the second DNN model and TabNet model achieve the lowest RMSE and MAE, indicating their superior predictive performance. .