LLM as Forecasting Planner: Training-Free Text Conditioning for Time-Series Foundation Models

Short summary

LAFP is a training-free framework that combines time-series foundation models (TSFMs) with LLMs for text-conditioned forecasting. The frozen TSFM proposes numerical trajectories while the LLM acts as policy and value function via Monte Carlo tree search, avoiding distortion of temporal structure. Experiments across two TSFM backbones and four LLMs show consistent improvements without retraining either model.

• •LAFP formulates text-conditioned forecasting as a planning problem over TSFM-generated trajectories
• •Uses MCTS with a Ranker LLM as policy and Judge LLM as value function—no retraining needed
• •Consistent improvements across Chronos/TimesFM backbones and four LLMs on Context-is-Key and Time-MMD benchmarks

Generated with AI, which can make mistakes.