Task-Aware Exchange Rate Forecasting: A Unified Empirical Framework for Level, Return, and Volatility Prediction

Exchange rate forecasting is a core issue in empirical finance, but in the majority of studies, the impact of target construction, feature representation, and model structure is not disaggregated, and the evaluation is confined to one prediction problem. The three tasks are: (1) prediction of $E{R}_{t+1}$; (2) prediction of $r_{t+1}=E{R}_{t+1}-E{R}_t$; and (3) prediction of future volatility defined as the sample standard deviation of $r_{t+1},\dots ,r_{t+10}$, where the horizon is the next 10 recorded observations. The seven model families are evaluated with a strict chronological 70%/15%/15% split; $ER$ is excluded as a direct raw predictor, deep models use seeds 7, 42, and 123, and task-specific financial benchmarks are included. The findings reveal markedly different out-of-sample forecasting behavior across the three tasks. For level prediction, the no-change random walk gives RMSE = 0.0005923 and $R^2=0.99815$, while Linear Regression gives RMSE = 0.0005939 and $R^2=0.99814$; the difference is not significant (DM = 0.480, $p=0.631$). For return prediction, Linear Regression gives RMSE = 0.0005920 and $R^2=-0.00061$ and does not significantly outperform the zero-return benchmark (DM = $-0.083$, $p=0.934$). For future volatility, the best average finance-aware learned model is FeatureAttention_Only (RMSE = $0.0004407\pm 0.0000180$; $R^2=-0.027\pm 0.085$), and its three-seed ensemble gives RMSE = 0.0004298 and $R^2=0.024$. FeatureAttention_Only improves squared-error loss relative to EWMA and GARCH in the fixed hold-out, but is not significantly superior to GARCH under QLIKE; rankings also vary across forecast origins.