Repackaging Temporal Evidence: A Unifying Interface for Temporal Prediction

Hao-Run Cai, Hao-Chen Liu, Ji-Yang Zhao, Yu-Jie Lai, Han-Jia Ye

Keywords: temporal prediction, irregular time series, temporal shift, tabular data, deep learning

Abstract: Temporal tabular prediction and irregular time-series prediction often use related longitudinal records, but they define different prediction protocols. Temporal tables expose a timestamped covariate row and test future-period generalization, whereas irregular time-series tasks expose entity-specific observations around a queried time. We study this difference as an evidence-packaging issue rather than a boundary between data modalities. Our unified protocol interface specifies the information available at query time before a predictor is applied. It identifies two failure modes in direct cross-domain transfer: entity collapse, where records from unrelated entities are treated as one sequence, and temporal relation mismatch, where a tabular row hides the source sequence needed by a time-series model. The same interface also yields protocol-level recoveries: context-based tabular predictors can be evaluated on irregular histories through entity contexts, and selected temporal tables can be recovered as entity histories or sequence-to-label examples.

Submission Number: 75