Improving Forecast Accuracy for Retail Supply Chains Using ARIMA and Machine Learning Approaches

In this paper, we focus on undertaking a systematic comparative study of four forecasting methods ARIMA, Gradient Boosting (GB), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM)—for forecasting the weekly demand of products in the retail industry. The models are tested using real-world time series data consisting of ten products and ten geographical areas under two scenarios: (i) forecasting demand across regions per product and (ii) forecasting products demand per region. Forecast accuracy evaluated using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Results show that BiLSTM was better than all other approaches, with RMSE and MAE reductions of up to 42.35% and 40.10%, respectively, when compared with ARIMA. These superior results are attributed to BiLSTM’s ability to capture complex temporal dependencies through its bidirectional architecture. LSTM and GB had modest improvements compared to ARIMA which underperformed, largely due to its limitations in modeling nonlinearity and non-stationarity. These results indicate the success of deep-learning methodology in retail forecasting and hint at attractive futures present for hybrid architectures and incorporation of exogenous variables.