Title: Graph-Augmented Retrieval for Cross-Entity Financial Sentiment Analysis: A Comparative Study

Abstract

While Retrieval-Augmented Generation (RAG) is essential for anchoring large language models in specialized datasets, traditional vector-based implementations struggle to capture the complex, multi-entity structures required for financial market analysis. This study offers a detailed comparison between a novel two-hop Graph-RAG architecture and a conventional vector-only baseline in the context of cross-entity financial sentiment analysis. Our approach involves building a sentiment-weighted knowledge graph comprising 59 equity entities, derived from 255 news articles focused on 10 major technology stocks. To enhance dense retrieval, we implement intensity-filtered graph traversal across INFLUENCES edges, enabling the system to uncover relational evidence that standard vector searches miss.

We assessed both systems using 100 grounded queries—split into 30 Direct and 70 Relational types—evaluating performance through semantic similarity, entity recall, RAGAS metrics, latency benchmarks, and ablation studies. The results demonstrate that Graph-RAG provides a statistically significant boost in entity recall (+6.4%, p < 0.001, per Wilcoxon signed-rank test) and generates notably more pertinent responses for intricate multi-entity inquiries (+11.7% increase in Answer Relevancy). These benefits were most pronounced in relational question categories, where relevancy improved by 16.1%. Importantly, these enhancements did not compromise answer quality, as evidenced by a negligible change in semantic similarity (delta = +0.001, Cohen's d = 0.078). Although the mean latency rose by 22.6%, this was counterbalanced by an 80% decrease in latency variance. Furthermore, an ablation study regarding the graph traversal intensity threshold identified an inverted-U relationship with answer quality, establishing tau = 0.5 as the optimal setting, surpassing the production default of tau = 0.7. These insights highlight the precision-for-coverage trade-off inherent in graph-augmented retrieval and offer practical architectural recommendations for developers creating RAG systems tailored to multi-entity financial analysis.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC