An Empirical Evaluation of COUTA Time-Series Anomaly Detection under Different Evaluation Protocols

Experimental conclusions in time-series anomaly detection depend not only on the model but also on threshold selection and evaluation protocols. Taking COUTA as the object, without changing the model structure, this research trains and tests on SMAP, MSL, and PSM separately, compares original point-wise evaluation, Point Adjustment (PA), and Train-Q99 alarm results in which the threshold is determined only by the 99th percentile of valid training scores, and also conducts sensitivity experiments on PSM at 5 synthetic training contamination levels. Every group of experiments employs five model random seeds and reports the mean and population standard deviation. Results of this paper show that PA-F1 on SMAP and MSL is 0.7547 and 0.8910, respectively, whereas Train-Q99 F1 is only 0.0130 and 0.0933; the Train-Q99 F1 on PSM is 0.4568, but the normal-point false positive rate reaches 0.1844. From 0% to 10% additional contamination, PSM’s Oracle Raw F1 decreases from 0.5476 to 0.4515, while PA-F1 increases from 0.8849 to 0.9470. This research establishes a traceable evaluation result set across multiple datasets and seeds, quantifies the gap between offline optimal evaluation, and fixed-threshold alarming, and reveals the directional divergence between point-wise evaluation and PA evaluation under training contamination. The results suggest that COUTA’s performance and contamination robustness are supposed to be interpreted jointly with raw scores, threshold information sources, and alarm burden. Differences in valid evaluation positions among protocols and preprocessing conditions are clarified in the main text.