Deep Learning and Statistical Approaches for Forecasting the Indonesian Rupiah Exchange Rate

Abstract

Accurate forecasting of exchange rates is essential for economic stability, investment strategy, and policy formulation. This study presents a comparative analysis of two distinct modeling approaches for predicting the Indonesian Rupiah (IDR) exchange rate against the US Dollar (USD): the Markov Switching Generalized Autoregressive Conditional Heteroskedasticity (MS-GARCH) model and the Long Short-Term Memory (LSTM) network enhanced with an attention mechanism. The MS-GARCH model captures volatility clustering and regime shifts, while the LSTM-Attention model learns complex nonlinear temporal dependencies. Using historical USD/IDR exchange rate data, both models are evaluated based on Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Empirical results show that the LSTM-Attention model achieves higher forecasting accuracy; however, the MS-GARCH model provides superior interpretability and insight into structural volatility. These findings underscore the importance of aligning model choice with forecasting objectives—highlighting that while deep learning offers enhanced predictive capability, statistical models remain valuable for risk analysis and financial diagnostics. The results support a complementary use of both methods in financial forecasting applications.