MARAG-Fin: An Intelligent Multi-agent RAG-LLM Architecture Integrating Financial News Sentiment and Time Series Data for Data-driven Trading Decision-making

Abstract

Financial markets present a highly dynamic and information-rich environment where effective decision-making requires both accurate forecasting and strong reasoning capabilities. Traditional quantitative models often struggle to capture unstructured knowledge from documents, while large language models (LLMs) excel in reasoning but lack reliable temporal forecasting. To address this gap, this research examines the effectiveness of a multi-agent framework based on Agentic Retrieval-Augmented Generation (RAG) for document-based reasoning, stock index prediction, and investment decision-making. The system combines LLM agents with contextual retrieval and a NeuralProphet-based dynamics agent for price forecasting, creating an adaptive framework suited to market dynamics. Evaluation focuses on three dimensions: reasoning quality, predictive accuracy, and backtested investment strategies on the S&P 500 and NASDAQ-100. Results show that ReAct Prompting outperformed Zero-shot Prompting, achieving an overall reasoning score of 0.92, while GPT-4.1 reached 0.94. The dynamics agent produced accurate forecasts with MAPE of 0.038-0.039. Backtests demonstrated that the S&P 500 strategy achieved a 19.85% return and 28.32% CAGR with stronger risk-adjusted performance, whereas the NASDAQ-100 delivered 57.19% returns with higher volatility. These findings highlight that integrating Agentic RAG, ReAct Prompting, and dynamics agents yields an adaptive and effective multi-agent system for investment decisions.