<![CDATA[We study monthly retail demand forecasting with an LSTM using a large real-world transaction history. The model is trained on historical monthly aggregates and evaluated on a two-month held-out horizon (January–February 2022) across 450 items. To avoid conflating optimization with forecasting performance, we report test-set accuracy using RMSE and MAPE, while training/validation losses are used only for model selection via early stopping. We run a controlled sweep over learning rate, batch size, and hidden units and find that the learning rate primarily drives convergence and accuracy, whereas the other two hyperparameters are secondary under our configuration. We also document practical data-preparation choices (outlier handling and chronological splitting) common in retail deployments. The study is framed as a single-customer case to ensure a consistent assortment and complete history; we therefore discuss external validity and reproducibility guidance. These findings help practitioners prioritize hyperparameter-tuning effort and set realistic expectations for short-horizon monthly demand forecasting in operational settings]]>
Copyrights © 2025
