FinKario: Event-Enhanced Automated Construction of Financial Knowledge Graph
A team presents FinKario, an auto updated knowledge graph from equity research, and FinKario-RAG, a two stage graph retriever for stock forecasting in China A shares. Backtests Aug 2024 to Mar 2025 show Sharpe 4.93, ARR 2.63, accuracy 0.581, beating LLMs and institutional strategies by 18.81 and 17.85 percent, and surpassing next best by 58 percent in Sharpe. Removing the Event Graph cuts performance over 80 percent. Short window and China scope limit claims noted.
What it examines
The paper tackles stale financial knowledge and long unstructured reports by introducing FinKario, a dynamic financial knowledge graph with attribute and event subgraphs built via template‑guided LLM extraction and quality control, and FinKario‑RAG, a two‑stage graph retrieval system, boosting stock trend prediction in backtests.
What it concludes
Experiments show higher returns, Sharpe, and accuracy than financial LLMs and institutional strategies; ablations confirm event knowledge and graph retrieval matter. Applications include equity research analysis, event extraction, RAG copilots, and strategy design. Limitations include modality gaps and evolving data; future work adds tables, charts, and time‑series.
Evidence objects
In China A-shares backtests (Aug 2024--Mar 2025), FinKario-RAG delivered Sharpe 4.93, ARR 2.63, accuracy 0.581; beat LLMs by 18.81% and institutional strategies 17.85%; topped runner-up Sharpe 58% and returns 31%.
key_findings bullet 1 · key_findings · validation V0
FinKario builds a dual graphAttribute for fundamentals, Event for time-sensitive driverstemplate-guided LLM extraction in CFA/JPM formats, Wisconsin causal analysis, FIBO; Tushare normalizes data; two-stage RAG targets entity/time, expands linked context.
key_findings bullet 2 · key_findings · validation V0
Dataset: 305,360 entities, 9,625 triples, 19 relations. Ablations: Event Graph essentialremoval slashes performance >80%. FinKario-RAG beats vanilla/graph-light RAG. Strengths: dynamic updates, schema-automation; limits: short window, China-only, LLM reliance, no costs/live.
key_findings bullet 3 · key_findings · validation V0
Introduces an event-enhanced financial knowledge graph from equity research and graph-based RAG for dynamic market reasoning, tackling knowledge lag and unstructured reports. Novelty lies in template-driven CFA/FIBO-guided schema, dual attribute/event subgraphs, real-time QC updates, and two-stage retrieval. Backtests reportedly beat financial LLMs and institutional strategies, though methodological originality is incremental.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … predictions (Lim1). To evaluate the effectiveness of FinKario, we conduct backtests comparing the predictive accuracy of our method with traditional financial LLMs and …
Source row: 845 · abstract type: snippet