TimeRAF: Retrieval-Augmented Foundation Model for Zero-Shot Time Series Forecasting

Time series forecasting plays a crucial role across numerous domains, driving rapid development in the field. With the advent of large models, time series foundation models (TSFMs) have exhibited great generalization capabilities, such as zero-shot learning, through large-scale pre-training. Meanwhile, Retrieval-Augmented Generation (RAG) methods are widely employed to enhance the performance of foundation models on unseen data across various domains, including Large Language Models (LLMs). To explore the integration of TSFMs with retrieval-augmented methods, we introduce TimeRAF, a Retrieval-Augmented Foundation model for zero shot time series Forecasting. A learnable retriever is employed and trained in an end-to-end fashion to extract useful information from a curated time series knowledge base. Additionally, we propose an approach called Channel Prompting for knowledge integration. Augmented by the retrieved knowledge, our TimeRAF demonstrates significant improvement across various domain and datasets. Furthermore, TimeRAF can leverage specialized knowledge bases to meet diverse application requirements. Extensive ablation studies and visualizations are provided to validate the effectiveness of our approach.