TSFMAudit: Data Contamination Auditing in Forecasting Time Series Foundation Models

2026-05-24

AI Summary

This study addresses the critical issue of pretraining data contamination in time series foundation models, which can lead to inflated performance evaluations due to overlap with benchmark test sets. To tackle this problem, the authors propose TSFMAudit, the first systematic approach for detecting such contamination by analyzing dynamic signatures during probe fine-tuning—specifically, faster loss convergence and reduced backbone parameter updates on contaminated samples. Departing from conventional discrete-text-based auditing paradigms, TSFMAudit introduces adaptation efficiency as a novel criterion for contamination detection. Extensive experiments across six prominent time series foundation models and 187 datasets demonstrate that TSFMAudit substantially outperforms ten baseline methods adapted from large language model auditing techniques, establishing a new standard for reliable evaluation in time series modeling.

Abstract

Time series foundation models (TSFMs) are increasingly pretrained on large corpora, raising concerns that evaluation datasets may have been exposed during pretraining and thus yield overly optimistic performance estimates. Auditing such contamination is challenging in time series because signals are continuous and heterogeneous, and often lack corpus documentation. To the best of our knowledge, this is the first work to study pretraining contamination auditing for TSFMs. We formalize the problem of pretraining contamination auditing for TSFMs and propose TSFMAudit, a method based on probe adaptation dynamics. Our key intuition is that contamination manifests as unusually efficient adaptation: after a fine tuning probe, contaminated datasets tend to exhibit faster loss reduction with smaller backbone movement. We evaluate TSFMAudit on 6 TSFMs and 187 datasets using documented training source evidence as supervision, and compare against 10 competitive baselines adapted from the LLM literature.

Problem

• data contamination
• time series foundation models
• pretraining auditing
• evaluation bias

Innovation

• data contamination auditing
• time series foundation models
• probe adaptation dynamics
• pretraining leakage
• model evaluation integrity