What is the right direction for time series anomaly detection benchmarking: evidence from evaluation of linear models

ICLR 2026 Conference Submission18474 Authors

Keywords: time series anomaly detection, benchmarking, deep learning, linear models

TL;DR: Simple linear regression can outperform deep learning in time series anomaly detection, highlighting the need for stronger baselines and more challenging benchmarks.

Abstract:

Time series anomaly detection (TSAD) progress has been accompanied by a persistent increase in architectural sophistication. In this work, we revisit this trend and demonstrate that a simple score based on a closed-form solution for an ordinary least squares (OLS) regression model outperforms state-of-the-art deep learning baselines. Through extensive evaluation on both univariate and multivariate TSAD benchmarks, we show that linear regression achieves superior accuracy and robustness while requiring orders of magnitude fewer resources.
Our further analysis identifies the types of anomalies that can and cannot be reliably captured by linear models, providing insights into their strengths and limitations. Overall findings indicate that current benchmarkings would benefit from inclusion of simple methods as well as more intricate problems that would do require deep learning-based solutions. Thus, future research should consistently include strong linear baselines and, more importantly, develop new benchmarks with richer temporal structures pinpointing the advantages of deep learning models.

Primary Area: learning on time series and dynamical systems

Submission Number: 18474