Adaptive Conformal Anomaly Detection for Time Series Monitoring

A novel method for post-hoc adaptive conformal anomaly detection in time series monitoring has been introduced, utilizing pre-trained foundation models without the need for fine-tuning. This approach generates interpretable anomaly scores presented as p-values, facilitating transparent decision-making. By employing weighted quantile conformal prediction bounds, it learns optimal weighting parameters from historical predictions, allowing for effective calibration amid distribution shifts and maintaining stable false alarm rates with out-of-sample assurances. As a model-agnostic solution, it seamlessly integrates with foundation models, enabling swift deployment in environments with limited resources. This innovation tackles industrial challenges such as insufficient data, a shortage of training expertise, and the demand for prompt inference.