REATS: LLM Reasoning-based Ensemble Learning for Adaptive Time Series Forecasting

2026-08-10

AI Summary

This work addresses the limitations of existing time series ensemble methods, which often rely on fixed rules or opaque black-box models and thus struggle to produce interpretable, sample-adaptive weight assignments. To overcome this, the paper introduces REATS, a novel framework that pioneers the integration of large language models (LLMs) into ensemble routing for time series forecasting. REATS fuses structured textual descriptions with numerical features and leverages chain-of-thought reasoning to generate interpretable and adaptive ensemble weights. The approach employs a hybrid input pipeline, retrieval-augmented priors, and a percentage-based tabular format, trained via a two-stage strategy combining supervised fine-tuning and GRPO reinforcement learning. Evaluated across eight benchmarks, REATS significantly outperforms current methods while offering natural language explanations, strong transferability, and robust out-of-domain generalization.

Abstract

Due to the diversity of real-world time series, no single forecasting model consistently dominates across all samples. Ensemble learning addresses this by combining complementary model strengths, yet existing methods rely on fixed rules or black-box models based solely on numerical inputs, failing to leverage LLM reasoning for interpretable weighting decisions. We propose REATS, which leverages LLM reasoning capabilities as an intelligent ensemble router that jointly processes textual temporal pattern descriptions and numerical features to produce interpretable, sample-adaptive ensemble weights through chain-of-thought reasoning. To enable effective LLM-based ensembling, we study its key design choices and propose: (i) a structured input pipeline that transforms raw time series into hybrid textual--numerical representations with fixed token cost, enabling rule-based chain-of-thought construction without API dependency, augmented with retrieved similar-sample priors; (ii) a diverse multi-row weight supervision scheme coupled with a token-efficient percentage-table format that reduces numerical complexity and mitigates LLM hallucinations; and (iii) a two-stage fine-tuning framework combining SFT with GRPO, where a reciprocal reward mapping transforms the continuous unbounded MSE gap into bounded signals with amplified near-oracle sensitivity, addressing the uniform sensitivity and outlier-dominated advantage compression inherent in naive reward designs for regression-based GRPO. Experiments on eight benchmarks demonstrate that REATS outperforms competitive ensemble baselines while providing natural language explanations and demonstrating strong transfer learning and out-of-domain generalization to unseen candidate models.

Problem

• time series forecasting
• ensemble learning
• large language models
• interpretable weighting
• adaptive forecasting

Innovation

• LLM reasoning
• ensemble learning
• time series forecasting
• chain-of-thought
• adaptive weighting