Time Series Forecasting via Reasoning: A Slow-Thinking Approach with Reinforcement Fine-Tuned LLMs

Accepted at ACM CIKM 2026 · Full Research Paper · Oral Presentation

Paper · CIKM 2026 Program · Code

📰 News

• [Aug. 2026] 🎉 Our paper Time-R1 has been accepted at ACM CIKM 2026 as a full research paper and is scheduled for an oral presentation!

📖 Abstract

Time-R1 investigates slow-thinking reasoning for time series forecasting. Unlike conventional forecasting methods that directly map historical observations to future values, Time-R1 trains large language models to explicitly analyze temporal dynamics before generating forecasts. Our framework combines supervised fine-tuning (SFT) for warm-up adaptation with reinforcement learning (RL) for reasoning optimization. It introduces a fine-grained multi-objective reward and GRIP (Group-based Relative Importance for Policy Optimization), which uses non-uniform sampling to encourage the exploration of effective reasoning paths. Experiments demonstrate competitive forecasting performance across diverse datasets.

This repository contains the official implementation of our CIKM 2026 paper:

Time Series Forecasting via Reasoning: A Slow-Thinking Approach with Reinforcement Fine-Tuned LLMs

🌟 Overview

Overview of Time-R1.

Large Language Models (LLMs) have demonstrated strong reasoning capabilities, but applying structured multi-step reasoning to time series forecasting remains challenging. Time-R1 adopts a two-stage reinforcement fine-tuning (RFT) framework with forecasting-oriented multi-objective rewards:

1. Stage 1 — SFT for Warm-up Adaptation: Teach the model foundational temporal analysis and structured reasoning using supervised fine-tuning on synthetic chain-of-thought data.
2. Stage 2 — RL for Reasoning Optimization: Optimize forecasting and reasoning trajectories using rewards informed by ground-truth alignment, multi-horizon accuracy, and domain-specific principles. GRIP uses non-uniform group sampling to encourage effective reasoning patterns.

GRIP optimization framework.

📚 Resources

• Training Dataset: Scripts and configurations for preparing training and evaluation datasets.
• Time-R1 Model: Code for training and inference under the two-stage RFT framework.
• Source Code: Training and evaluation implementations in this repository.

⚙️ Key Features

• Slow-Thinking Time Series Reasoning: Enables deliberate, step-by-step analysis of temporal patterns before forecasting.
• Two-Stage RFT Framework: Combines SFT warm-up adaptation with RL-based reasoning optimization.
• GRIP Optimization: Uses group-based relative importance and non-uniform sampling to explore effective reasoning paths.
• Fine-Grained Multi-Objective Rewards: Incorporates forecasting accuracy and temporal reasoning objectives into policy optimization.
• Forecasting Performance: Demonstrates competitive performance across diverse real-world datasets.

🚀 Quick Start

Installation

We recommend Python 3.10+ in a clean conda environment. The original implementation specifies CUDA ≥ 12.4 and cuDNN ≥ 9.8.0 for training and inference.

git clone https://github.com/ustc-time-series/Time-R1.git
cd Time-R1
conda create -n time-r1 python=3.10
conda activate time-r1

# Install the verl framework and dependencies
pip install --no-deps -e .
pip install -r requirements.txt

🛠️ Training

Time-R1 RL Training

bash scripts/time-r1.sh

📈 Evaluation

cd eval
python main.py

📝 Citation

If you find Time-R1 useful for your research, please cite our CIKM 2026 paper:

@inproceedings{zhou2026timer1,
  title     = {Time Series Forecasting via Reasoning: A Slow-Thinking Approach with Reinforcement Fine-Tuned LLMs},
  author    = {Zhou, Yitong and Luo, Yucong and Cheng, Mingyue and Liu, Qi and Wang, Jiahao and Wang, Daoyu and Chen, Enhong},
  booktitle = {Proceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM '26)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2506.10630}
}

The citation will be updated with official ACM publication metadata once available.