Counterfactual Necessity: Sufficiency-Based Time-Series Explanations Are Only Half the Story

arXiv 2607.21573·medium signal

Hongnan Ma, Yiwei Shi and Mengyue Yang argue that faithful explanations of time-series classifiers must identify subsequences that are not only sufficient to preserve the black-box prediction but necessary to it, adding a counterfactual necessity criterion to standard sufficiency-based attribution. A subsequence can be sufficient while being entirely redundant — the model would predict the same thing without it. For teams shipping explainability on forecasting or anomaly-detection models, sufficiency-only attributions can confidently highlight the wrong window.