Causal Analysis Framework Evaluates Time Series Foundation Models.

The shift towards time series foundation models, moving from one-to-one to one-to-many application relationships, introduces significant concentration risk. Many critical forecasting applications could be exposed to the same biases and failure modes of a single model. While this centralization offers economies of scale in development, robust validation before deployment becomes paramount. This research proposes a causal analysis framework specifically designed to investigate the ability of time series foundation models to preserve underlying time series patterns. The methodology involves intervening on parameterized synthetic time series generators and then observing the corresponding changes in model output under controlled conditions. This allows for a systematic identification of biases and potential failure points. Applying this framework to Chronos-2 and TimesFM-2.5 across six distinct time series patterns, the study found safe configurations for trend and harmonic oscillation. However, it also revealed a bias in both models towards overestimating persistence, sudden failures for both models with regime switch patterns, and a failure for TimesFM-2.5 with the energy-release pattern. These findings are potentially attributable to the models' pretraining data, highlighting the importance of such pre-deployment analysis.