Did We Actually Fix It? An Independent Adversarial Stress-Test of Post-Point-Adjustment Evaluation Metrics for Time-Series Anomaly Detection

This feed aggregates new ML/statistics arXiv papers (15 Jul 2026). Most security-relevant is an independent adversarial stress-test of post-point-adjustment metrics for time-series anomaly detection (TSAD): the authors evaluate 12 adopted metrics across 250 UCR Anomaly Archive series and five additional benchmarks using trivial and adversarial no-skill score generators. Key findings: affiliation-F1 and ROC-based metrics (including VUS-ROC) are widely gameable by no-skill detectors (high gameability rates across benchmarks), while PR-based metrics and PA%K are substantially more robust. The kit