Latent Inference-Time Guidance of Time Series Foundation Models
Abstract
Time Series Foundation Models (TSFMs) currently provide state-of-the-art results in forecasting tasks. They are available out-of-the-box and rely on in-context learning to make their predictions, which makes the quality of their performance highly sensitive to the user-selected lookback, covariates, horizon and training data distributions. In practise, the quality of the forecasts are variable but complementary, which highlights the need for a principled ensembling approach, rather than selecting the best context. This paper introduces Latent Inference-Time Guidance for TSFMs, which adaptively combines a pool of TSFM forecasts through a time-dependent latent space with independent components. The framework comes equipped with identifiability and reconstruction guarantees, whilst maintaining the off-the-shelf aspect of foundation models. We provide experiments on datasets at various frequencies and from multiple domains: these show that the approach is competitive with traditional ensembling approaches.
Related
- Steepest Guidance: A Practical and Principled Approach to Inference-Time Alignment of Flow and Diffusion-based Models
- Distillation of Synthetic Data for Time Series Foundation Models
- Cadence: Error-Bounded Lossy Compression of Demand Time Series with a Time-Series Foundation Model
- SGA: Uncertainty Quantification for Multi-Step Forecasting in Time Series Foundation Models
- Latent-driven Imputation and Forecasting for partially observed Time Series (LIFTS)
BibTeX
@misc{hashimotocullen2026latent,
title = {{Latent Inference-Time Guidance of Time Series Foundation Models}},
author = {Chloé Hashimoto-Cullen and Amaury Durand and Laurent Bozzi and Benjamin Guedj and Yannig Goude and Sylvain Le Corff},
howpublished = {HAL},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.38058},
eprint = {2609.38058},
archivePrefix = {arXiv},
url = {https://hal.science/hal-05773148v1},
note = {CC BY-NC-SA 4.0},
}
RIS
TY - UNPB
TI - Latent Inference-Time Guidance of Time Series Foundation Models
AU - Chloé Hashimoto-Cullen
AU - Amaury Durand
AU - Laurent Bozzi
AU - Benjamin Guedj
AU - Yannig Goude
AU - Sylvain Le Corff
PY - 2026
DA - 2026/09/29
PB - HAL
DO - 10.48550/arxiv.2609.38058
UR - https://hal.science/hal-05773148v1
AB - Time Series Foundation Models (TSFMs) currently provide state-of-the-art results in forecasting tasks. They are available out-of-the-box and rely on in-context learning to make their predictions, which makes the quality of their performance highly sensitive to the user-selected lookback, covariates, horizon and training data distributions. In practise, the quality of the forecasts are variable but complementary, which highlights the need for a principled ensembling approach, rather than selecting the best context. This paper introduces Latent Inference-Time Guidance for TSFMs, which adaptively combines a pool of TSFM forecasts through a time-dependent latent space with independent components. The framework comes equipped with identifiability and reconstruction guarantees, whilst maintaining the off-the-shelf aspect of foundation models. We provide experiments on datasets at various frequencies and from multiple domains: these show that the approach is competitive with traditional ensembling approaches.
ER -