TimeRFT: Stimulating Generalizable Time Series Forecasting for TSFMs via Reinforcement Finetuning

2026/05/31 by Siyang Li, Yize Chen, Zijie Zhu +4
Engineering · Computer Science · #eess.SP #cs.AI #cs.CV #cs.LG

paper · pdf

15 pages, 8 figures, In Submission

arxiv created 2026/07/31 · arxiv updated 2026/08/03

Abstract

Time Series Foundation Models (TSFMs) have demonstrated strong generalization capability and data efficiency in time series forecasting through large-scale pretraining. However, adapting TSFMs to downstream forecasting tasks remains challenging due to temporal distribution shifts and varying data availability. Specifically, the non-stationary and uncertain nature of time series data leads to discrepancies between historical training and future forecasting distributions, making existing Supervised FineTuning (SFT)-based adaptation vulnerable to overfitting and limited generalization. Moreover, forecasting tasks often operate under varying data regimes, requiring TSFMs to extract generalizable temporal patterns from limited training samples. To address these challenges, we propose Time series Reinforcement FineTuning (TimeRFT), a reinforcement learning-based adaptation paradigm for TSFMs. TimeRFT introduces two forecasting-oriented training recipes: (i) A quality-aware temporal reward mechanism providing fine-grained credit assignment by holistically evaluating the contribution of each prediction step to overall forecasting performance. (ii) A difficulty-aware data selection strategy prioritizing informative time series samples with generalizable forecasting patterns. Extensive experiments on diverse real-world forecasting benchmarks demonstrate that TimeRFT consistently surpasses SFT-based adaptation methods across various real-world forecasting tasks with different data regimes, achieving improved prediction accuracy and enhanced generalization against unforeseen distribution shifts.

Citations

• Quadratic Direct Forecast for Training Multi-Step Time-Series Forecast Models
• Enhancing Time Series Forecasting through Selective Representation Spaces: A Patch Perspective
• On the Generalization of SFT: A Reinforcement Learning Perspective with Reward Rectification
• Multi-Scale Finetuning for Encoder-based Time Series Foundation Models
• Time Series Forecasting as Reasoning: A Slow-Thinking Approach with Reinforced LLMs
• MIRA: Medical Time Series Foundation Model for Real-World Health Data
• Improving Data Efficiency for LLM Reinforcement Fine-tuning Through Difficulty-targeted Online Data Selection and Rollout Replay
• TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning
• What Can RL Bring to VLA Generalization? An Empirical Study
• BLAST: Balanced Sampling Time Series Corpus for Universal Forecasting Models
• This Time is Different: An Observability Perspective on Time Series Foundation Models
• GRPO-LEAD: A Difficulty-Aware Reinforcement Learning Approach for Concise Mathematical Reasoning in Language Models
• Efficient Reinforcement Finetuning via Adaptive Curriculum Learning
• Understanding R1-Zero-Like Training: A Critical Perspective
• Empowering Time Series Analysis with Synthetic Data: A Survey and Outlook in the Era of Foundation Models
• LIMR: Less is More for RL Scaling
• Process Reinforcement through Implicit Rewards
• Sundial: A Family of Highly Capable Time Series Foundation Models
• TimeHF: Billion-Scale Time Series Models Guided by Human Feedback
• Kimi k1.5: Scaling Reinforcement Learning with LLMs
• ChatTS: Aligning Time Series with LLMs via Synthetic Data for Enhanced Understanding and Reasoning
• Tulu 3: Pushing Frontiers in Open Language Model Post-Training
• On Designing Effective RL Reward at Training Time for LLM Reasoning
• Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Experts
• GIFT-Eval: A Benchmark For General Time Series Forecasting Model Evaluation
• Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning
• VinePPO: Refining Credit Assignment in RL Training of LLMs
• Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts
• Qwen2.5-Coder Technical Report
• Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvement
• Beyond LoRA: Exploring Efficient Fine-Tuning Techniques for Time Series Foundational Models
• VisionTS: Visual Masked Autoencoders Are Free-Lunch Zero-Shot Time Series Forecasters
• Sequence to Sequence Reward Modeling: Improving RLHF by Language Feedback
• PowerPM: Foundation Model for Power Systems
• Deep Time Series Models: A Comprehensive Survey and Benchmark
• Foundation Models for Time Series Analysis: A Tutorial and Survey
• Teaching Large Language Models to Reason with Reinforcement Learning
• MOMENT: A Family of Open Time-series Foundation Models
• FreDF: Learning to Forecast in the Frequency Domain
• Unified Training of Universal Time Series Forecasting Transformers
• DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence
• ReFT: Reasoning with Reinforced Fine-Tuning
• Let's Verify Step by Step
• Dish-TS: A General Paradigm for Alleviating Distribution Shift in Time Series Forecasting
• A Time Series is Worth 64 Words: Long-term Forecasting with Transformers