Time series models answer associations, not interventions, study finds

A new study published on arXiv examines the capabilities of covariate-aware time series foundation models (TSFMs) in answering "what-if" questions. The research found that models like Chronos-2, TiRex-2, TabPFN-TS, and TimesFM-2.5, when trained on observational data, tend to provide answers based on associations rather than true interventions. These models exhibit a "confounding gap," responding more influenced by confounding factors than standard regression techniques. While the models' internal representations can decode data-generating regimes and true effects, their forecasts do not reflect this understanding. The study suggests that interventional post-training can significantly reduce this gap, with fine-tuning methods showing promise in improving performance on interventional benchmarks. AI

IMPACT This research highlights limitations in current time series models for causal inference, suggesting a need for interventional training to improve their reliability in "what-if" scenarios.

RANK_REASON The cluster contains a research paper detailing findings about the behavior of time series foundation models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• Chronos-2 Forecasting Model
• DagsHub
• Gotit.pub
• Hugging Face
• ScienceCast
• TabPFN-TS
• TimesFM-2.5
• TiRex-2

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →