Enhancing Capabilities of Forecasting Models Using LLMs and Agentic RAG

Time series forecasting is critical across domains like finance, energy, and meteorology, yet traditional models struggle with data sparsity, complex spatio-temporal dependencies, and generalization to unseen datasets. This paper proposes a novel framework that integrates Large Language Models (LLMs) with an agentic Retrieval-Augmented Generation (RAG) architecture to enhance forecasting capabilities. Our approach employs a hierarchical multi-agent system, dynamic retrieval with Transformer-based embeddings, and attention-based feature fusion to process time-series data, unstructured text, and external knowledge. The framework achieves superior accuracy, scalability, and zero-shot generalization by augmenting pre-trained LLMs with real-time context, addressing limitations in computational efficiency and cross-domain adaptability. Evaluated on benchmark datasets (e.g., M4, GluonTS, Weather), our model outperforms state-of-the-art baselines, including SARIMAX, XGBoost and LSTM. This work demonstrates the potential of agentic RAG to transform time series forecasting, offering a scalable, domain-agnostic solution for real-world applications.