A Modified ADAM Optimizer for Enhancing LSTM Performance in Predicting Prodigo Shoe Sales

Abstract

This study aims to improve the accuracy of sales prediction for Prodigo shoe products by modifying the ADAM optimizer in a Long Short-Term Memory (LSTM) model. The main challenges include unstable sales trends and limited training data. The ADAM optimizer is modified by preserving gradient direction to improve weight update stability and convergence. The dataset consists of 387 weekly sales records from January 2018 to June 2025, and data augmentation is performed using TimeGAN to generate synthetic data. The methodology includes Min-Max normalization, sequence generation using sliding window, and hyperparameter optimization via grid search. Model performance is evaluated using Mean Absolute Error (MAE) and Dynamic Time Warping (DTW). The results show that the modified ADAM outperforms the standard ADAM in both scenarios, with and without synthetic data. The best model, using synthetic data, achieves an MAE of 290 and a DTW of 3 with more efficient training time. These findings indicate that combining modified ADAM with data augmentation improves temporal pattern learning and prediction accuracy, and the model has been implemented as a Flask-based web service.