Abstract

Accurate retail sales forecasting requires models that can handle linear seasonality, non-linear dynamics, and consumer sentiment at the same time. Seasonal AutoRegressive Integrated Moving Average (SARIMA) captures periodic trends well but cannot learn non-linear residual patterns. Long Short-Term Memory (LSTM) networks have the opposite problem: they lack any mechanism for incorporating sentiment-based exogenous inputs. Prior hybrid SARIMA-LSTM work exists, but none of it has tested the systematic integration of Valence Aware Dictionary and Sentiment Reasoner (VADER) sentiment polarity, as a statistically validated exogenous variable, within a bidirectional residual-correction framework on retail transactional data. This study addresses this limitation through the development of a hybrid forecasting framework. The proposed framework has three components: Seasonal AutoRegressive Integrated Moving Average with eXogenous variables (SARIMAX) captures linear trends, seasonality, and sentiment-driven influences. A Bidirectional LSTM (Bi-LSTM) network trained on SARIMAX residuals models the non-linear temporal dependencies that SARIMAX misses. Prophet then handles additional trend-shift correction. VADER scores derived from customer product reviews serve as the exogenous regressors. The framework is evaluated on the Super Store Sales dataset covering 2011 to 2014, which contains 51,290 transactions. Stationarity was confirmed via the Augmented Dickey-Fuller (ADF) test (ADF statistic: −4.8676, p < 0.0001). A formal ablation study isolates each component’s contribution. The full hybrid model achieves a Root Mean Squared Error (RMSE) of 61,588.96, Mean Absolute Error (MAE) of 49,270, and Mean Absolute Percentage Error (MAPE) of 9.82%, outperforming SARIMA (RMSE: 75,936.21), Prophet (RMSE: 74,365), and Neural-Basis Expansion Analysis for Time Series (N-BEATS) (RMSE: 65,300); the sentiment-augmented SARIMAX baseline reaches 70,544.74. The Diebold-Mariano test confirms statistical superiority over both SARIMAX (p < 0.01) and Prophet (p < 0.05). Sentiment integration alone cuts RMSE by 7.57%; Bi-LSTM residual modelling adds a further 14.54%, with practical implications for inventory management and dynamic pricing.