From Recoverability to Functional Use: Auditing Temporal Reports in Time-Series Forecasting

Time-series forecasters increasingly accompany numerical predictions with explicit temporal reports, such as delays or selected history, but a correct report need not describe the information actually used by the forecast. We separate this problem into three questions: whether the target temporal structure is recoverable from the observed trajectory, whether the model reports it correctly, and whether the forecast functionally depends on the reported history. For point delays, we show that accurate prediction does not imply correct temporal use: two candidate delays can become statistically distinguishable as the trajectory grows even when substituting one for the other incurs little per-step prediction error. On controlled time-series tasks, learned forecasters can remain primarily dependent on history outside the reported structure even when the target is recoverable, the report is correct, and the forecast is accurate. Finally, report-controlled routing substantially improves report--forecast alignment without an observed loss in predictive accuracy. These results motivate separate statistical and functional validation of temporal reports.