Reproduce “Latent Inference-Time Guidance of Time Series Foundation Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chloé Hashimoto-Cullen, Amaury Durand, Laurent Bozzi, Benjamin Guedj, Yannig Goude, Sylvain Le Corff
LITiG-TSFM combines forecasts from frozen time-series foundation models with a learned, time-varying latent guidance layer, rather than fine-tuning models. The layer can mix or nonlinearly correct forecasts using observations and covariates. In a Washington Bicycle Share ablation, nine strong experts were augmented with one, three, or five weak experts; LITiG-TSFM remained robust and competitive, but experts were ranked on test segments.
Key insights
- A learned, time-varying latent state can provide nonlinear guidance beyond fixed or linear forecast weights, while operating only on foundation-model outputs.
- Theoretical guarantees concern denoised-signal identifiability and reconstruction under stated assumptions; they do not establish full latent-component identification or universal forecasting accuracy.
- In tests with weak experts added, LITiG-TSFM was comparatively stable, suggesting it could limit their influence in that setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.