Comparing LSTM, Bi-LSTM, and Transformer Across Feature Scenarios for Rupiah Exchange Rate Forecasting

Abstract

This study compares LSTM, Bi-LSTM, and Transformer models for one-step-ahead JISDOR forecasting across eight feature scenarios using historical JISDOR, IHSG, Brent oil prices, and inflation. The dataset comprised 1,300 chronological observations from January 2021 to May 2026. Evaluation used chronology-safe preprocessing, four expanding-window walk-forward folds, validation-based model selection, five stochastic seeds, a naive persistence benchmark, statistical testing, and Integrated Gradients. Among the deep-learning models, LSTM under S1 achieved the lowest mean MAPE of 0.6662%, followed by Bi-LSTM at 0.6899% and Transformer at 1.8135%. However, naïve persistence achieved a lower mean MAPE of 0.2618%, and all model–scenario combinations were significantly worse after Holm-adjusted Diebold–Mariano testing. External variables did not improve forecasting accuracy over historical JISDOR alone. Integrated Gradients showed that historical JISDOR received the largest attribution in all three architectures. These results highlight the importance of temporal validation and simple benchmarks in exchange-rate forecasting.