REATS Uses LLM Reasoning for Adaptive Time Series Forecasting

Time series forecasting is challenging because no single model consistently performs best across all diverse real-world data. Ensemble learning, which combines multiple models, offers a solution by leveraging their complementary strengths. However, current ensemble methods typically rely on fixed rules or black-box models that only consider numerical inputs, missing the opportunity to use the advanced reasoning capabilities of Large Language Models (LLMs) for more interpretable and adaptive weighting decisions. This research introduces REATS (LLM Reasoning-based Ensemble Learning for Adaptive Time Series Forecasting). REATS positions an LLM as an intelligent ensemble router. This LLM jointly processes both textual descriptions of temporal patterns and traditional numerical features. Through chain-of-thought reasoning, it generates interpretable, sample-adaptive ensemble weights, effectively deciding which base models to prioritize for a given forecast. To make LLM-based ensembling effective, REATS incorporates several key design choices: a structured input pipeline that converts raw time series into hybrid textual-numerical representations with controlled token costs, augmented with retrieved similar-sample priors; a diverse multi-row weight supervision scheme using a token-efficient percentage-table format to reduce numerical complexity and mitigate LLM hallucinations; and a two-stage fine-tuning framework combining Supervised Fine-Tuning (SFT) with Generative Reinforcement Learning from Policy Optimization (GRPO), using a reciprocal reward mapping for better sensitivity. Experiments on eight benchmarks show REATS surpasses competitive ensemble baselines, offers natural language explanations, and demonstrates strong transfer learning and out-of-domain generalization.