Does Cross-Domain Pre-Training Truly Help Time-Series Foundation Models?

Inspired by the success of pre-training large language models, recent efforts have explored cross-domain pre-training for time-series foundation models (TSFMs). However, the distinct data generation dynamics and contextual limitations of time-series data challenge the direct transferability of LLM strategies to TSFMs. In this paper, we investigate whether cross-domain pre-training truly benefits TSFMs. Through systematic experiments, we reveal that while cross-domain pre-training can enhance performance in certain domains, it may also cause severe negative transfer in others due to domain disparities in sampling frequencies and evolution patterns. Surprisingly, transfer effects are often counterintuitive: unrelated domains can yield significant gains, whereas related domains may induce degradation. These findings highlight the need for tailored pre-training strategies that address the unique characteristics of time-series data. Our study provides actionable insights to guide the development of more effective TSFMs.