Do Macroeconomic Variables Improve Rupiah Forecasts under Conditional Heteroskedasticity?

Abstract

Rupiah exchange-rate forecasting studies are usually univariate and rarely evaluated under a leakage-free protocol against a strong naive benchmark. This study tests whether three macroeconomic variables (Federal Funds Rate, Bank Indonesia policy rate, and Indonesian Consumer Price Index) improve forecasts from a linear ARIMA-GARCH model, whose GARCH component captures conditional heteroskedasticity. The resulting ARIMAX-GARCH is compared with a nested ARIMA-GARCH control and random walks with and without drift. Using daily JISDOR data (2018 to 2026), models are estimated by quasi-maximum likelihood in 1,000-day rolling windows at horizons of one day to one quarter. Point forecasts are evaluated by MAE, RMSE, and MAPE with Clark-West and Giacomini-White tests, and interval forecasts by coverage, the interval score, and the Kupiec and Christoffersen tests. The macroeconomic variables add no predictive value over the nested control at any horizon under either test. This null result persists under an error-correction extension, Student-t errors, alternative window sizes, EGARCH and GJR-GARCH variances, and two separate subperiods, consistent with the Meese-Rogoff puzzle within the daily linear specification tested. The model beats the random walk without drift, but only because of the depreciation trend, since the random walk with drift reproduces about 97% of the gain. The GARCH component adds value in interval rather than point forecasting, giving well-calibrated 95% intervals about 12% narrower than constant-variance benchmarks at one day. Macroeconomic variables therefore do not improve daily rupiah forecasts from ARIMA-GARCH models, and gains over a random walk should be tested against a drift benchmark.