Do Time-Series Forecasters Use the Right History: Recoverability, Recovery, and Functional Use of Temporal Delays

2026-08-10

AI Summary

Although existing time series forecasting models often report accurate lags, it remains unclear whether their predictions genuinely rely on the historical information they claim to use. This work proposes an evaluation framework based on recoverability analysis, integrating matched masking tests with lag-report verification to systematically assess the actual historical inputs leveraged by a model. Our investigation reveals that in models such as N-HiTS and TCN, over half of the cases—reaching up to 92.7%—exhibit correct reported lags and near-optimal predictive performance yet fail to meaningfully utilize the claimed historical context. To address this discrepancy, we introduce a hard one-hot routing mechanism that enforces alignment between the reported lag and the actual computational path, thereby enabling precise and controllable modeling of temporal dependencies.

Abstract

Forecast accuracy does not tell us which past inputs produced a prediction. We separate three questions for time-series models with known delay structure: can the true delay be recovered from the observed data, does the model report it, and does the forecast actually use the same history? We first derive input-conditioned recoverability measures that separate intrinsic ambiguity from model error. We then prove that a delay report can become arbitrarily reliable while forecast risk approaches the oracle even though the predictor still uses the wrong lag. This failure also appears in finite samples on the point-delay task: among forecasts with a correct delay report and normalized excess risk within 10\% of the oracle, the reported history is functionally unused under our matched masking test in 55.4\% of N-HiTS cases and 92.7\% of TCN cases. Finally, we show that routing the prediction through the reported history removes off-report bypass paths; a hard one-hot control achieves exact fixed-report alignment. The main conclusion is simple: a good forecast, even with a correct delay report, does not show that the model used the right history.

Problem

• time-series forecasting
• temporal delays
• recoverability
• history usage
• delay reporting

Innovation

• recoverability
• temporal delay
• matched masking
• forecast alignment
• time-series forecasting