Anatomi Pergerakan Nilai Tukar di Era Ketidakpastian: Pengujian Efisiensi Pasar dan Proses Martingale Kurs KRW/IDR Melalui Regresi Dinamis ARIMA versus Ensembel Machine Learning

Sari Desi Ratna
Sahrudin Sahrudin

Prosiding Seminar Nasional Official Statistics
Vol 2026 No 1 (2026): Seminar Nasional Official Statistics 2026

The Korean Won (KRW)/Indonesian Rupiah (IDR) exchange rate is important for trade, investment, and remittances, but its non-stationary, near-random-walk behavior complicates forecasting. This study compares seven statistical, regression, and machine learning models using 2,651 daily observations from December 2018 to April 2026: Naïve, ETS, ARIMA(0,1,2)(0,0,2)[7], deterministic trend with ARIMA errors, stochastic trend, XGBoost, and Random Forest. On a 90- day hold-out set, deterministic trend with ARIMA errors achieved the lowest RMSE (0.147) and remained best across six rolling-origin windows. However, Diebold–Mariano tests found no significant differences from Naïve, ARIMA, stochastic trend, and ETS (p=0.138–0.314). In contrast, all statistical/regression models significantly outperformed XGBoost and Random Forest (MASE > 4). The findings support the Efficient Market Hypothesis and the importance of multi-period evaluation and formal significance testing.