MODEL HYBRID LSTM-XGBOOST UNTUK PREDIKSI NILAI TUKAR RUPIAH TERHADAP DOLAR AMERIKA

Abstrak

Exchange rate prediction is crucial for economic stability, investment, and international trade. The high volatility of the rupiah against the US dollar makes accurate prediction challenging. This study proposes a hybrid Long Short-Term Memory (LSTM) and eXtreme Gradient Boosting (XGBoost) model to improve exchange rate prediction accuracy. The LSTM captures long-term temporal dependencies in time-series data, while XGBoost models the residuals from LSTM predictions to capture non-linear patterns and macroeconomic influences. The dataset spans January 2020 to December 2024, including daily Rupiah exchange rates, BI 7-Day Reverse Repo Rate, inflation, and US Dollar Index (DXY). Results demonstrate that the hybrid model provides relatively better performance compared to individual LSTM and XGBoost models, which experienced significant prediction failures. While the hybrid approach shows potential for short-term forecasting with acceptable margin error, the model's ability to explain data variability remains limited. This indicates that although the hybrid method offers improvement over individual models, substantial enhancements in model architecture and data preprocessing are still required for practical applications.

Keywords: Long Short-Term Memory, eXtreme Gradient Boosting, Machine Learning, Prediction, Exchange Rate