TSFMAudit: Data Contamination Auditing in Forecasting Time Series Foundation Models

In three linesTSFMAudit, first contamination auditing method for time series foundation models (TSFMs). Detects whether evaluation datasets were exposed during pretraining by analyzing fine-tuning adaptation dynamics: contaminated data exhibits unusually fast loss reduction. Evaluated on 6 TSFMs and 187 datasets.

Summary generated by Claude — human-verified