TSFMAudit: Data Contamination Auditing in Forecasting Time Series Foundation Models

TSFMAudit：时间序列基础模型预测中的数据污染审计

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.

摘要： 时间序列基础模型（TSFMs）越来越多地在大型语料库上进行预训练，这引发了人们的担忧：评估数据集可能在预训练过程中已被泄露，从而导致过于乐观的性能评估。在时间序列领域审计此类污染具有挑战性，因为时间序列信号具有连续性和异构性，且往往缺乏语料库文档记录。

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.

据我们所知，这是首个研究 TSFM 预训练污染审计的工作。我们对 TSFM 的预训练污染审计问题进行了形式化定义，并提出了 TSFMAudit，这是一种基于探测适应动态（probe adaptation dynamics）的方法。我们的核心直觉是：污染表现为异常高效的适应性——在进行微调探测后，受污染的数据集往往表现出更快的损失下降，且模型主干（backbone）的变化更小。

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.

我们使用有据可查的训练源证据作为监督，在 6 个 TSFM 模型和 187 个数据集上评估了 TSFMAudit，并与从大语言模型（LLM）文献中改编的 10 个竞争基线进行了对比。