Skip to content

Adaptive Conformal Anomaly Detection with Time Series Foundation Models for Signal Monitoring

Home / Papers / Adaptive Conformal Anomaly Detection with Time Series Foundation Models for Signal Monitoring

Adaptive Conformal Anomaly Detection with Time Series Foundation Models for Signal Monitoring

Authors: Natalia Martinez Gil, Fearghal O’Donncha, Wesley M. Gifford, Nianjun Zhou, Dhaval C. Patel, Roman Vaculín Date: 2026-04-22 Paper ID: openalex:2604.20122

Summary

This paper presents a post-hoc, model-agnostic conformal anomaly detection method designed for time series monitoring using pre-trained foundation models. By applying weighted quantile conformal prediction, the method provides statistically rigorous, interpretable p-values for anomaly detection while ensuring robust false alarm control. The adaptive mechanism learns weighting parameters from historical predictions, allowing the system to maintain performance under non-stationary conditions without the need for additional fine-tuning.

Key Contributions

  • Introduces a post-hoc adaptive conformal anomaly detection framework that integrates with pre-trained time series foundation models without requiring fine-tuning.
  • Utilizes weighted quantile conformal prediction to maintain false alarm rate control and provide statistically interpretable p-values for anomalies.
  • Implements an adaptive weighting mechanism for calibration that effectively handles distribution shifts while remaining computationally efficient for resource-constrained environments.

Open Questions & Future Work

Key Concepts

  • weighted-conformal-anomaly-detection: A conformal prediction framework that uses adaptively learned weights on historical quantiles to maintain rigorous false alarm control under distribution shifts.

Archivist Review

I have approved the core methodological concept of using weighted conformal prediction for adaptive anomaly detection, as it provides a robust, reusable framework for calibration in non-stationary time series. The open question was refined to capture the fundamental tension between statistical coverage guarantees and non-exchangeable temporal data streams, moving away from the paper-specific phrasing.

Approved Concepts

  • Weighted Conformal Anomaly Detection: Combines conformal prediction with adaptive weighting for non-stationary anomaly detection, which is a powerful mechanism for robust signal monitoring in time series.

Approved Open Questions

  • Conformal Anomaly Detection Bottleneck: This represents a foundational bottleneck in deploying rigorous statistical monitoring tools in high-stakes time-series environments where temporal dependencies and distribution shifts are the norm.

Metadata & Links

url
https://arxiv.org/abs/2604.20122
paper_id
2604.20122
paper_source
openalex
domain
time-series
tags
anomaly-detectionconformal-predictionfoundation-modelstime-series-forecasting
architectures
datasets
concept_slugs
weighted-conformal-anomaly-detection
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-04-25T05:37:23Z