An Evaluation of Ensemble Strategies for Time Series Anomaly Detection

Real-time anomaly detection in time series is a critical task in domains where timely identification of unexpected events can prevent failures or losses. A popular approach relies on forecasting models that generate prediction intervals, with anomalies flagged when new observations fall outside expected ranges. In this study, we investigate two ensemble strategies for anomaly detection in the Numenta Anomaly Benchmark dataset: interval ensembling, which aggregates prediction intervals across multiple models before anomaly detection, and voting ensembling, which aggregates binary anomaly decisions after detection. Using classical univariate forecasting models as base models, we evaluate twelve different ensemble configurations. Results show that interval ensembling consistently outperforms voting ensembling in terms of both Accuracy and Fl score, with statistically significant differences confirmed by statistical tests. The findings demonstrate that leveraging predictive uncertainty directly provides a more robust and reliable basis for surveillance-style anomaly detection than relying on post-hoc consensus of anomaly flags.