Evaluating Time-Series Foundation Model Reliability Under Structural Change in African Currency Markets

Time-series foundation models (TSFMs) are increasingly deployed zero-shot in markets that are poorly represented in their pretraining data. We ask whether information available before the forecast horizon can identify when a TSFM should not be trusted. Using 2,848 forecast origins across 34 African USD exchange-rate series and three models (Chronos-Bolt, TimesFM 2.5, and Moirai 2.0), we find no consistent point-forecast advantage over persistence. Under documented IMF AREAER de-facto regime reclassifications, however, nominal 80\% prediction intervals cover only 54.7–68.6\% of post-reclassification outcomes while their widths barely adjust. Outside regime-change episodes, some forecast failures were predictable in advance: recent price staleness helped identify risky point forecasts, while model interval width and limited prior local performance helped identify unreliable probabilistic forecasts. These signals do not transfer reliably to regime-transition origins. A macrofinancial extension does not close this gap for point-forecast risk, although broad money relative to reserves provides secondary evidence for probabilistic transition risk. Overall, TSFM reliability is partially diagnosable under ordinary conditions, but structural regime change remains a blind spot for uncertainty calibration.