The Emergence of Automated Financial Knowledge Graphs

Why financial KGs matter

Financial data is notoriously complex. Corporate disclosures, market news and research reports come in long, unstructured text; events evolve quickly; and relationships between entities change over time. Traditional databases capture numerical indicators but often miss relational patterns and context. Financial knowledge graphs address this gap by representing companies, sectors, indicators and events as nodes connected by edges, enabling multi, hop reasoning and explainable analytics. Automated KG construction is critical because manual curation is too slow to keep up with the pace of new filings and market events.

Event Enhanced knowledge graphs with FinKario

In August 2025 the FinKario project introduced one of the first event enhanced financial KGs. FinKario automatically extracts information from equity research reports using prompt based templates designed by investment professionals. The resulting KG contains over 305,000 entities, 9,625 relational triples and 19 distinct relation types. A key innovation is that FinKario integrates real time company fundamentals and market events so that the Knowledge Graph can incorporate earnings releases, product launches and regulatory changes as soon as they happen. A two stage FinKario RAG retrieval strategy delivers relevant subgraphs to LLMs, improving both context and efficiency. Back testing experiments show that FinKario with FinKario RAG achieved 18.81% better stock trend prediction accuracy than financial LLMs and 17.85% better performance than institutional strategies.

Dynamic knowledge graphs and FinDKG

While FinKario focuses on static research reports, FinDKG (July 2024) explores dynamic knowledge graphs for capturing evolving market trends. FinDKG uses a fine‑tuned large language model, the Integrated Contextual Knowledge Graph Generator (ICKG), to extract entities and relationships from a large corpus of financial news articles. The result is an open‑source dynamic knowledge graph that can track events over time. To analyse the time varying graph, the authors proposed the KGTransformer, an attention based graph neural network that learns representations of entities while incorporating temporal dynamics. Experiments demonstrate that KGTransformer improves link prediction metrics and, when applied to thematic investing it outperforms existing thematic exchange traded funds.

Building blocks of dynamic knowledge graphs

FinDKG formalizes a dynamic knowledge graph as a set of temporal quadruples. The ICKG model uses engineered prompts to extract these quadruples from news text, assembling them into an event indexed graph. The KGTransformer then leverages meta entity types and graph attention mechanisms to learn representations that capture both structure and time.

Key takeaways

1. Automated financial KGs like FinKario integrate real‑time events and company fundamentals, providing structured insights that improve stock trend prediction.

2. FinDKG introduces a dynamic, time indexed knowledge graph extracted from news articles with a fine‑tuned LLM and analyses it using graph neural networks.

3. The combination of prompt driven extraction and graph based retrieval/representation enables knowledge graphs that can keep pace with market volatility and support real‑time analytics.