New LAFP framework uses LLM as planner for text-conditioned time-series forecasting without retraining
Combines a frozen TSFM with LLM-guided MCTS to better predict from context.
Time-series forecasting typically ignores rich context like events or constraints expressed in natural language. A new paper by Nguyen et al. from VinAI introduces LLM as Forecasting Planner (LAFP), which treats forecasting as a planning problem over trajectories generated by a frozen time-series foundation model (TSFM). The TSFM acts as a simulator that proposes numerical continuations, while two LLM roles—Ranker (policy) and Judge (value function)—guide a Monte Carlo tree search to select the best candidate trajectory. All components are frozen, requiring no retraining or fine-tuning.
Evaluated on the Context-is-Key and Time-MMD datasets using Chronos and TimesFM backbones with four different LLMs (including GPT-4 and open-source models), LAFP consistently outperformed baselines across all model combinations. The approach enables text-conditioned forecasts that account for events the raw time series cannot reveal, without the distortion caused by asking an LLM to directly modify numerical values. This training-free, plug-and-play strategy could significantly simplify building context-aware forecasting systems for finance, supply chain, and climate applications.
- LAFP uses a frozen TSFM as simulator and LLM as policy/value function via Monte Carlo tree search, no retraining needed.
- Tested on Context-is-Key and Time-MMD datasets with Chronos and TimesFM backbones and four LLMs (including GPT-4).
- Consistent improvements over baselines, bridging text and time-series without modifying model weights.
Why It Matters
Enables context-aware time-series forecasting without expensive retraining, useful for finance, logistics, and climate predictions.