Comparison of Hyperparameter Optimization Methods for LSTM-Based XAU/USD Forecasting and Web-Based System Implementation

Abstract

Unequal search budgets and single-seed experiments can confound comparisons of hyperparameter optimisation methods in financial forecasting. This study compares grid search, random search, and Bayesian optimisation for tuning a long short-term memory model to forecast the XAU/USD closing price for the next trading day. The dataset comprised 1,705 daily observations from January 2020 to July 2026 using open, high, low, close, and release-date-aligned United States inflation. Each method evaluated the same 32 configurations using three random seeds, resulting in 96 candidate-model evaluations per method. Performance was assessed on 37 independent testing dates and descriptively examined using a five-fold post-selection walk-forward diagnostic without repeating hyperparameter optimisation within each fold. All methods selected the same configuration and produced a mean testing MAPE of 2.785052% and an ensemble MAPE of 2.696161%. Grid Search reached the final-best configuration earlier, but naïve persistence achieved the lowest MAPE of 1.358325%. Thus, optimisation improved the LSTM relative to the predefined baseline but did not outperform persistence. The procedures were also implemented in a Streamlit application. The findings are limited to the examined dataset, search space, seeds, testing period, and computational environment.