Latent Inference-Time Guidance of Time Series Foundation Models

2026-09-29

AI Summary

This study addresses the challenge that the prediction quality of time series foundation models is highly context-dependent and lacks principled ensemble methods. To overcome this, we propose a latent inference-time guidance framework that leverages independent component analysis to construct a temporally dependent latent space with identifiability and reconstruction guarantees. This approach adaptively fuses predictions from multiple models, effectively integrating complementary information without requiring optimal context selection while preserving out-of-the-box usability. Extensive evaluations across multi-domain datasets demonstrate that the proposed method achieves performance comparable to conventional ensemble techniques while exhibiting superior robustness. Consequently, this work establishes a novel paradigm for the efficient ensembling 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.

Problem

• Time Series Foundation Models
• Forecasting
• Ensembling
• Context Sensitivity

Innovation

• Time Series Foundation Models
• Latent Inference-Time Guidance
• Ensembling
• Identifiability
• In-context Learning