Impact-Driven Event Covariates for Time-Series Foundation Models

Pretrained time-series foundation models can condition on future-known covariates, but event covariates are often limited to binary indicators or sparse categorical labels. This is problematic for forecast-based anomaly detection: in financial transaction series, recurring holidays, promotions, and settlement cycles induce predictable shifts that event-unaware systems may flag as false positives. We propose impact-driven event covariates, learned from event-centered residual responses across heterogeneous series rather than from event identity alone. Supplying these covariates to zero-shot Chronos-2 on a synthetic benchmark calibrated to empirical financial-transaction signatures reduces WAPE from 0.2872 to 0.2265 and improves anomaly-detection F1 from 0.2573 to 0.4580 relative to a no-event Chronos baseline. The learned covariates also transfer to unseen accounts with limited history without retraining. The results suggest that event covariates are most useful when their temporal alignment matches the underlying business process, providing an effective interface between domain event structure and pretrained time-series foundation models.