TimeRFT: Reinforcement Finetuning Boosts Time Series Foundation Models

A new paradigm called TimeRFT (Time series Reinforcement Finetuning) has been introduced to improve the adaptability of Time Series Foundation Models (TSFMs) for downstream forecasting tasks. TSFMs, which leverage large-scale pretraining for generalization and data efficiency, often struggle when fine-tuned on specific tasks due to temporal distribution shifts between training and testing data, leading to overfitting in supervised methods. Additionally, varying data availability across tasks challenges generalization. TimeRFT addresses these issues with two task-specific training recipes: a forecasting quality-based temporal reward mechanism that evaluates multiple facets of the forecast, and a reinforcement learning framework that optimizes the model under diverse data regimes. The approach aims to enhance robustness and generalization without relying solely on supervised fine-tuning.