ReAugment: Targeted Few-Shot Time Series Augmentation via Model Zoo-Guided Reinforcement Learning

Few-shot time series forecasting suffers from severe overfitting due to limited high-quality training data. We introduce ReAugment, a reinforcement learning (RL) framework that explicitly learns where and how to augment time series data. ReAugment maintains a zoo of forecasting models and measures prediction diversity across them to identify training samples that are most prone to model overfitting. These samples are "bottlenecks" for generalization and are used as anchor points in the augmentation process. We then employ an RL approach to learn data transformation policies, using a model zooguided reward function to bias the transformed data to overfit-prone regions of the training distribution that are most beneficial for generalization. A key advantage of the RL formulation is that it avoids backpropagating gradients through the forecasting models, thereby mitigating gradient vanishing. Empirical results across various benchmarks show that ReAugment consistently improves forecasting accuracy in both few-shot and standard settings. Code available at https: //github.com/ironllen/ReAugment .