Investigating simple target-covariate relationships for Chronos-2 and TabPFN-TS
Investigating simple target-covariate relationships for Chronos-2 and TabPFN-TS
Authors: Gaspard Berthelier, Mariia Baranova, Andrei-Tiberiu Pantea, Etienne Le Naour, Adrien Petralia, Tahar Nabil, Themis Palpanas Date: 2026-05-12 Paper ID: openalex:2605.12200
Summary
This paper investigates the ability of state-of-the-art Time Series Foundation Models (TSFMs), specifically Chronos-2 and TabPFN-TS, to capture fundamental target-covariate relationships. Through controlled diagnostic experiments, the authors reveal that while both models excel in broader benchmarks, TabPFN-TS significantly outperforms Chronos-2 in capturing these dependencies, particularly for short-term forecasting. The findings highlight that general zero-shot performance does not necessarily equate to robust modeling of explicit feature-target interactions.
Key Contributions
- Conducts controlled diagnostic experiments to evaluate the target-covariate integration capabilities of Chronos-2 and TabPFN-TS.
- Demonstrates that TabPFN-TS outperforms Chronos-2 in modeling simple target-covariate dependencies, particularly over short forecasting horizons.
- Challenges the assumption that superior general benchmark performance in TSFMs guarantees effective modeling of fundamental covariate-target relationships.
Open Questions & Future Work
Key Concepts
- tabpfn-ts: A Transformer-based foundation model for tabular and time series data that treats forecasting as an in-context regression problem.
Archivist Review
The analysis correctly highlights the contrast between generative TSFMs and in-context learners. I approved TabPFN-TS as a distinct architectural paradigm for time series, but rejected Chronos-2 as it is a specific model instance rather than a generic category. The open question is well-framed and reflects a significant ongoing challenge in the field.
Approved Concepts
- TabPFN-TS: The paper identifies it as a distinct model class for covariate-target relationship modeling in TSFMs compared to autoregressive foundations.
Approved Open Questions
- Unified Temporal-Tabular Foundation Architectures: Current time series foundation models typically excel at either univariate temporal modeling or feature-target regression, creating a performance gap in domains where both patterns are crucial for accuracy.
Rejected Candidates
- [concept] Chronos-2 (
chronos-2) - paper_local: Chronos-2 is a specific model implementation rather than a foundational methodological concept.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.12200
- paper_id
- 2605.12200
- paper_source
- openalex
- domain
- time-series
- tags
- forecastingtime-series-foundation-modelsbenchmarking
- architectures
-
- datasets
-
- concept_slugs
- tabpfn-ts
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-15T06:15:29Z