Well-Calibrated Until It Matters: EventRisk-Bench for Conditional Reliability of Time-Series Foundation Models

A forecasting model can benefit from event information without producing reliable uncertainty estimates during those events. EventRisk-Bench evaluates this distinction around SEC-acceptance-aligned Item 2.02 results disclosures for U.S. equities. We forecast the log of a daily variance proxy formed from squared hourly returns; its first component may span overnight. We compare Chronos-2, TimesFM-3, and Moirai-1.1-R-Base using identical 512-session histories and controlled event inputs. Across 1,987 test events and 50,000 sampled ordinary origins, probability calibration error is $0.004$–$0.009$ when pooled over the eligible weighted population, but $0.085$–$0.394$ on events. Event conditioning reduces weighted interval score by $21.7\%$ for Chronos-2 and $53.3\%$ for TimesFM-3, yet all three have worse event scores than a rolling Event-HAR baseline with the same history. Event-specific validation recalibration reduces event calibration error in all three models; only Chronos-2 meets every prespecified repair criterion. Follow-up matching and synthetic experiments support the distinction between useful event information and reliable uncertainty. Aggregate calibration and better event scores are insufficient evidence of event reliability.