FinLLM-KG: Knowledge Graph-Enhanced LLMs for Financial Sentiment Analysis

01 ABSTRACT

Existing financial sentiment LLMs often predict each article in isolation as black boxes, limiting contextual understanding and deployment in high-stakes decisions. The authors propose FinLLM-KG, which injects company-level knowledge from a financial knowledge graph at inference time as a natural-language profile or a few ranked factual triples, without updating model weights. Simulations show consistent portfolio performance improvement over a state-of-the-art baseline, with sentence- and fact-level attribution.

02 KEY FINDINGS

1. FinLLM-KG retrieves company-level facts from a financial knowledge graph built on SEC 10-K filings and injects them into causal LLMs at inference as a natural-language profile or ranked triples, without fine-tuning.
2. The framework attributes each article-level sentiment score to individual sentences and retrieved knowledge-graph facts, enhancing explainability and supporting human-in-the-loop review.
3. In simulations, FinLLM-KG consistently improves portfolio performance over the state-of-the-art baseline.