Title: DAD4TS: A Data-Augmentation-Driven Diffusion Model for Enhancing Time-Series Forecasting with Limited Data

Abstract: The scarcity of data poses a significant challenge in time-series forecasting. While data augmentation offers a viable solution, existing techniques often struggle to produce meaningful synthetic samples. To overcome this bottleneck, we introduce DAD4TS, a novel data augmentation framework leveraging diffusion models integrated with reinforcement learning, specifically tailored for time-series forecasting in low-data scenarios. Within the DAD4TS architecture, a data generator is co-trained alongside a time-series predictive model and governed by a reinforcement learning agent. This collaborative training process ensures the efficient generation of samples that directly enhance the forecast accuracy of the target model. To accommodate small-scale datasets, we replace standard Variational Autoencoder (VAE) approaches with mathematical projection techniques that map time-series data into geometric space for training the diffusion model. We evaluated the efficacy of DAD4TS through both qualitative and quantitative experiments, comparing it against seven baseline methods across eight time-series models and six real-world datasets. The results confirm the robustness and effectiveness of DAD4TS on five of these datasets.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC