Attenuated in-context identification in time-series foundation models: diagnosis under counterfactual inputs and repair by synthetic forced-system fine-tuning
Time-series foundation models attenuate counterfactual dynamics; synthetic forced-system fine-tuning partly restores response magnitude.
The paper tests whether covariate-aware time-series foundation models can answer what-if questions on forced systems with exact counterfactuals, comparing Chronos-2, TimesFM-2.5, and TabPFN-TS with classical identification. TimesFM-2.5 and TabPFN-TS are memoryless, while Chronos-2 identifies dynamics but attenuates them to 0.33–0.80 of the true effect. A 26-minute fine-tune on synthetic forced systems restores sensitivity to 0.83–0.96, yet classical identification stays clearly better on three of four measured plants and the fine-tune costs some univariate forecasting skill.
- TimesFM-2.5 and TabPFN-TS act memoryless through default covariate interfaces
- Chronos-2 predicts only 0.33–0.80 of the true counterfactual effect
- A 26-minute synthetic fine-tune restores sensitivity to 0.83–0.96
- Classical identification remains better on three of four measured plants
- AI models
- Chronos-2TimesFM-2.5TabPFN-TS
From the original article
1 min read at arxiv.orgCovariate-aware time-series foundation models (TSFMs) promise training-free what-if answers for instrumented plants: the change in output that a different future input would cause. We test this on forced engineering systems with exact counterfactuals, comparing Chronos-2, TimesFM-2. 5 and TabPFN-TS with classical system identification fitted to the same context.
By Hong-In Won · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.