Multifractal-State-Augmented Machine Learning for One-Trading-Day-Ahead USD/KRW Absolute-Return Forecasting

Abstract

Daily exchange-rate returns show weak average persistence, yet their multiscale organization may inform future movement magnitude. This study evaluates whether time-varying multifractal states improve one-trading-day-ahead forecasts of USD/KRW absolute returns. The raw dataset contains 6520 Federal Reserve H.10 observations from 1999 to 2024. After initialization of the 252-day rolling MFDFA window, 6268 feature-aligned observations remain for model estimation and evaluation. Seven scaling features augment Extra Trees, and forecasts are routed through pre-test terciles of singularity-spectrum width. Shuffled, phase-randomized, and IAAFT surrogates, matched-feature ablations, temporal-shift placebos, specification variants, and expanding-window thresholds assess robustness. In the locked 2020–2024 test of 1249 observations, the routed model yields an RMSE of 0.367 and an MAE of 0.271, improving on market-only Extra Trees by 1.627% and 2.245%, respectively. Temporal displacement produces upper-tail p-values of 0.020 for RMSE and 0.010 for MAE. The model’s R2 is 0.045 and its forecast–realization correlation is 0.218. Validation-calibrated 95% and 99% bands reduce interval scores by 2.160% and 5.730%, although coverage remains below nominal. External-currency gains are mixed and do not survive multiplicity adjustment. The results indicate modest, time-aligned, USD/KRW-specific predictive information; they do not establish universal transferability, causal effects, directional prediction, or trading profitability.