Time Series Sales Forecasting: A Hybrid Deep Learning Regularization Approach

Sales forecasting is essential in diverse sectors, driving informed decision-making and efficient resource allocation. In this research, we concentrate on enhancing retail sales forecasting, specifically within the furniture industry, utilizing deep neural network models. Drawing insights from an accessible dataset containing historical sales data of a global superstore, we investigate various forecasting techniques, including classical time-series methods like SARIMA, with advanced neural network models like RNN and bidirectional LSTM (BiLSTM), to capture temporal dependencies effectively. We propose a hybrid CNNBiLSTM model with regularization in BiLSTM layers to improve generalization and forecasting accuracy. Our study evaluates performance using MAE, RMSE, and $R^{2}$ score as error metrics. The results illustrate our proposed approach’s superiority with an MAE of 1.25, RMSE of 1.56, and an $R^{2}$ score of 0.53, demonstrating significant improvements in forecasting accuracy. Moreover, the introduction of regularization underscores our model’s performance, refining its predictive capabilities. This study not only unveils the potential of harnessing deep neural networks for precise retail sales prediction but also emphasizes the pivotal role of normalization techniques in nurturing robust forecasting models, holding significant implications for informed decision-making in the dynamic retail landscape and beyond.