Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Comparative Study of Adaptation Strategies for Time Series Foundation Models in Anomaly Detection

Miseon Park, Kijung Yoon

TL;DR

This paper tackles time series anomaly detection (TSAD) by evaluating Time Series Foundation Models (TSFMs) pretrained on large, heterogeneous data as universal backbones. It systematically compares zero-shot inference, full fine-tuning, and parameter-efficient fine-tuning (PEFT) strategies across multiple TSAD benchmarks, demonstrating that adapted TSFMs outperform task-specific baselines on metrics like $AUC$-$PR$ and $VUS$-$PR$, especially under class imbalance. The study finds that PEFT methods such as $LoRA$, $OFT$, and $HRA$ often match or exceed full fine-tuning while reducing computational cost, though MoE-based TSFMs may require more specialized PEFT to close the gap. Overall, TSFMs with PEFT emerge as a scalable, efficient approach for TSAD, with clear guidance on architecture-dependent adaptation and directions for future work in multivariate and streaming anomaly detection.

Abstract

Time series anomaly detection is essential for the reliable operation of complex systems, but most existing methods require extensive task-specific training. We explore whether time series foundation models (TSFMs), pretrained on large heterogeneous data, can serve as universal backbones for anomaly detection. Through systematic experiments across multiple benchmarks, we compare zero-shot inference, full model adaptation, and parameter-efficient fine-tuning (PEFT) strategies. Our results demonstrate that TSFMs outperform task-specific baselines, achieving notable gains in AUC-PR and VUS-PR, particularly under severe class imbalance. Moreover, PEFT methods such as LoRA, OFT, and HRA not only reduce computational cost but also match or surpass full fine-tuning in most cases, indicating that TSFMs can be efficiently adapted for anomaly detection, even when pretrained for forecasting. These findings position TSFMs as promising general-purpose models for scalable and efficient time series anomaly detection.

A Comparative Study of Adaptation Strategies for Time Series Foundation Models in Anomaly Detection

TL;DR

This paper tackles time series anomaly detection (TSAD) by evaluating Time Series Foundation Models (TSFMs) pretrained on large, heterogeneous data as universal backbones. It systematically compares zero-shot inference, full fine-tuning, and parameter-efficient fine-tuning (PEFT) strategies across multiple TSAD benchmarks, demonstrating that adapted TSFMs outperform task-specific baselines on metrics like - and -, especially under class imbalance. The study finds that PEFT methods such as , , and often match or exceed full fine-tuning while reducing computational cost, though MoE-based TSFMs may require more specialized PEFT to close the gap. Overall, TSFMs with PEFT emerge as a scalable, efficient approach for TSAD, with clear guidance on architecture-dependent adaptation and directions for future work in multivariate and streaming anomaly detection.

Abstract

Time series anomaly detection is essential for the reliable operation of complex systems, but most existing methods require extensive task-specific training. We explore whether time series foundation models (TSFMs), pretrained on large heterogeneous data, can serve as universal backbones for anomaly detection. Through systematic experiments across multiple benchmarks, we compare zero-shot inference, full model adaptation, and parameter-efficient fine-tuning (PEFT) strategies. Our results demonstrate that TSFMs outperform task-specific baselines, achieving notable gains in AUC-PR and VUS-PR, particularly under severe class imbalance. Moreover, PEFT methods such as LoRA, OFT, and HRA not only reduce computational cost but also match or surpass full fine-tuning in most cases, indicating that TSFMs can be efficiently adapted for anomaly detection, even when pretrained for forecasting. These findings position TSFMs as promising general-purpose models for scalable and efficient time series anomaly detection.
Paper Structure (19 sections, 2 figures, 4 tables)

This paper contains 19 sections, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Time Series Foundation Models
  4. Parameter-Efficient Fine-Tuning
  5. TSFMs for Anomaly Detection
  6. Methods
  7. Dataset
  8. TSFM Backbones
  9. Zero-shot inference.
  10. Fine-tuning and parameter-efficient adaptation.
  11. Adaptation Strategies and Training Protocol
  12. Evaluation Metrics
  13. Baseline Anomaly Detection Methods
  14. Experimental Result
  15. Conventional TSAD Baselines
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Relationship between model size and zero-shot anomaly detection performance. Mean VUS-PR is plotted against the number of model parameters for different TSFM variants evaluated on the TSB-AD-U benchmark. Bubble size reflects relative model capacity.
  • Figure 2: Comparison of VUS-PR and VUS-ROC performance for zero-shot, full fine-tuning, and PEFT methods on Moirai (left), Chronos (middle), and Time-MoE (right).