Performance Comparison of XGBoost and LSTM for Rupiah–USD Exchange Rate Prediction

Abstract

Fluctuations in the Rupiah exchange rate against the US Dollar (USD/IDR) pose challenges to economic stability, making accurate prediction important for policymakers, businesses, and investors. This study compares XGBoost and Long Short-Term Memory (LSTM) models in predicting the USD/IDR exchange rate using daily selling and buying rates from Bank Indonesia covering January 2010 to December 2025. The preprocessing stages include log return transformation, feature engineering using lag returns, rolling volatility, and moving average returns, as well as normalization using StandardScaler. Hyperparameter tuning was performed using random search, while model evaluation was conducted using walk-forward validation with periodic retraining. Model performance was measured using MAE, RMSE, SMAPE, and Directional Accuracy (DA). The results show that LSTM achieves better performance in MAE (41.41 IDR), SMAPE (0.2588%), and DA (52.74%), while XGBoost obtains a lower RMSE (56.63), indicating more stable performance in handling larger prediction errors. Overall, both models produce comparable predictive performance, with LSTM showing a slight advantage in average prediction accuracy and directional movement prediction