Time-Series Models May Not Use Reported Historical Data.

When time-series models generate predictions, it's often assumed that if they report using a specific historical input, they actually do. This paper challenges that assumption, showing that high forecast accuracy and even correct reporting of temporal delays do not guarantee the model genuinely utilizes that reported history. The researchers identify three key questions: whether the true delay can be identified from data, if the model reports it, and if the forecast truly uses it.The study introduces input-conditioned recoverability measures to distinguish between inherent data ambiguity and model errors. It demonstrates that a model can achieve near-oracle forecast risk and reliable delay reports while still functionally ignoring the reported lag. Experiments with models like N-HiTS and TCN show that in a significant percentage of cases, the reported history is unused. The paper concludes by suggesting that explicitly routing predictions through the reported history can enforce alignment, preventing "off-report bypass paths" and ensuring the model's internal logic matches its stated dependencies.