Hybrid LSTM-MLP Framework for Enhanced Sales Forecasting

Abstract

This study introduces a hybrid framework combining Long Short-Term Memory (LSTM) networks and Multi-Layer Perceptrons (MLP) to enhance sales forecasting through advanced feature processing. The LSTM component is employed to capture temporal dependencies in sequential sales data, while the MLP focuses on processing numerical and categorical features to complement the predictive accuracy. This integration leverages the strengths of both architectures, enabling more comprehensive and effective feature processing compared to traditional approaches. The methodology encompasses data preprocessing, model development, and evaluation across performance metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and the coefficient of determination (R2). By employing real-world sales datasets, the framework undergoes rigorous testing to compare its effectiveness against existing models. Experimental results show that the hybrid LSTM-MLP model achieves notable improvements, including reduced MAE, MSE, and RMSE, while delivering 99% R2 values, higher than standalone LSTM, MLP, and other traditional methods. These findings underscore the model’s potential to offer accurate and stable predictions, demonstrating its utility for forecasting tasks where tabular and sequential data interplay. The proposed framework not only contributes to advancing predictive methodologies in sales forecasting but also serves as a reference for adopting hybrid strategies in similar studies. Its effectiveness in handling feature complexity and producing reliable forecasts highlights its applicability in real-world forecasting scenarios and lays the foundation for further research in hybrid neural network designs.