FinLLM-KG: Knowledge Graph-Enhanced Large Language Model for Financial Sentiment Analysis in Algorithmic Trading

The study proposes a model-agnostic FinLLM-KG framework that combines causal large language models with financial knowledge graphs, injecting time-anchored company-level knowledge from SEC 10-K filings during inference without updating model weights. The framework can provide context as natural language company profiles or a small number of relevance-ranked factual triples, with almost no additional computational cost, while supporting attribution of sentiment predictions to specific sentences and knowledge graph facts, improving interpretability. Simulation results show that FinLLM-KG consistently improves portfolio performance compared to existing optimal baselines.