A Survey on Current Practices for Time Series Foundation Model Evaluation
Time series forecasting is present across several domains, such as finance, healthcare, energy, and operations, and foundation models promise to replace the traditional one-model-per-series workflow with a single zero-shot pretrained forecaster applied to these domains. Whether they deliver, and where they stop working, remains unknown. Following a survey (conducted on 1 June 2026 from Scopus) reported in line with the PRISMA guidelines, this study surveys 45 articles across five survey questions, covering what time series foundation models are and how they are designed and trained, how well they perform, what is known about their theoretical learnability limits, what phase diagrams and scaling laws reveal about them, and how controlled synthetic temporal environments are constructed and used to evaluate these models. A taxonomy of how models are built and how they are evaluated shows that the literature concentrates on building and benchmarking time series foundation models. Causal analysis of time series foundation models' abilities and limitations remains underexplored. We conclude that there is significant opportunity for future research in broadening the evaluation methods applied to time series foundation models.
Authors
- Marcos Machado (ORCID: https://orcid.org/0000-0003-1056-2368)
- Faizan Ahmed (ORCID: https://orcid.org/0000-0002-2760-6892)
- Mathis Jander (ORCID: https://orcid.org/0009-0004-9321-862X)
- Ties Dilling
Institutions
- Saxion (NL)
- University of Twente (NL)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24279624
- Primary Topic
- Time Series Analysis and Forecasting
- Type
- preprint
Funders
- European Commission