<![CDATA[This study asks whether the attribution of a predictive model of the Rupiah (IDR/USD) agrees with an independent structural decomposition, separately within volatility regimes. Hybrid econometric and machine-learning studies are evaluated almost exclusively by forecast error, leaving untested whether the learning stage weights the drivers the structural stage identifies. The pipeline combines a Structural Vector Autoregression (SVAR) with Cholesky identification, four-state volatility Markov-switching labeling, regime-conditioned features, walk-forward XGBoost, and SHAP attribution reconciled per regime with the Forecast Error Variance Decomposition (FEVD), on daily data from January 2010 to June 2026. The FEVD shows own-shock dominance of 92.5 percent at the 22-day horizon, robust to alternative orderings and to identification through heteroskedasticity (82.0 percent), though sub-period estimates range from 78.8 to 95.3 percent. On direction, XGBoost attains 56.4 percent at t+1 (Pesaran-Timmermann, p less than 0.001); on magnitude it does not beat a random walk at any horizon, whereas a GRU baseline does, so the predictive claim is confined to direction and gradient boosting is retained for its exact tree attribution. Exact permutation tests over all relabelings of the driver axis find agreement significant in only one of twelve regime-horizon cells, the tail regime at 22 days (cosine 0.902, p = 0.036), carried by 2020, while the own-dynamics share rises monotonically from 18.1 to 42.4 percent across regimes, significant at every horizon. The contribution is a falsifiable per-regime attribution audit with an explicit account of where it fails.]]>
