Frédéric Montet and Benjamin Pasquier (iCoSys, HEIA-FR/HES-SO) presented new research at ITISE 2025 in Granada on how today’s “foundation models” stack up for time-series forecasting in real-world settings. The work was supported by a young researcher grant from HES-SO won by Prof. Beat Wolf.

The study compares popular models from Amazon, Salesforce and Google with classic statistical and deep-learning methods, and tests them in three everyday situations: no training data (zero-shot), a little data (few-shot) and full training. The team also looked at speed and ease of use—what actually matters when deploying models in industry. In short: compact foundation models can often match or beat traditional approaches with minimal tuning, especially when data is scarce, while larger models bring extra accuracy at a higher compute cost. preprints202507.0279.v1 preprints202507.0279.v1

A pre-print of the paper, Benchmarking Foundation Models for Time-Series Forecasting, is available on Preprints.org (DOI: 10.20944/preprints202507.0279.v1).