New study finds common time-series anomaly detection metrics are gameable

By PulseAugur Editorial · [2 sources] · 2026-07-12 22:33

A new paper critically examines evaluation metrics for time-series anomaly detection (TSAD), finding that many commonly used metrics are susceptible to being gamed by simple, no-skill score generators. The research tested 12 metrics across multiple benchmarks, revealing that affiliation-F1 and ROC-based metrics like VUS-ROC are particularly vulnerable, while PR-based metrics and PA%K show more resilience. The authors release a stress-test harness and recommend prioritizing PR-based metrics or PA%K, treating affiliation-F1 and ROC-AUC variants with caution, and verifying metric performance on a per-benchmark basis. AI

IMPACT Highlights critical flaws in standard evaluation methods, potentially impacting the reliability of future research in anomaly detection.

RANK_REASON The cluster contains an academic paper detailing new research findings on evaluation metrics.

Read on arXiv stat.ML →

paper
other

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

New study finds common time-series anomaly detection metrics are gameable

COVERAGE [2]

arXiv stat.ML TIER_1 English(EN) · Zongye Lyu · 2026-07-15 04:00

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

arXiv:2607.11969v1 Announce Type: new Abstract: Point-adjustment (PA), long the default scoring protocol in time-series anomaly detection (TSAD), was shown by Kim et al. (2022) to award near-perfect F1 to random scores. The field migrated to replacement metrics: PA%K, range-based…
arXiv stat.ML TIER_1 English(EN) · Zongye Lyu · 2026-07-12 22:33

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

Point-adjustment (PA), long the default scoring protocol in time-series anomaly detection (TSAD), was shown by Kim et al. (2022) to award near-perfect F1 to random scores. The field migrated to replacement metrics: PA%K, range-based precision/recall, affiliation precision/recall,…

COVERAGE [2]

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

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

RELATED ENTITIES

RELATED TOPICS