FinKario: Event-Enhanced Automated Construction of Financial Knowledge Graph

In one sentence

Summary

The paper addresses the challenge of integrating rapidly evolving market events and unstructured equity research reports into LLM-based financial analysis, where static knowledge bases and simple chunking fail to provide timely, context-aware insights. The authors propose FinKario, a dual-structured financial knowledge graph comprising an attribute subgraph for stable fundamentals and an event subgraph for time-sensitive drivers, constructed automatically via prompt-driven extraction guided by professional institutional templates like the CFA handbook and FIBO ontology. This approach overcomes the limitations of manual schema engineering and slow update cycles, enabling dynamic, real-time knowledge population from raw markdown reports. The methodology includes a comprehensive pipeline for corpus acquisition, schema construction, knowledge population, and quality control refinement using Tushare data for attribute completion. To leverage this graph, the authors develop FinKario-RAG, a two-stage retrieval strategy that first identifies coarse-grained entities (stocks, dates) and then expands to fine-grained related entities and relationships to form a semantically coherent subgraph for reasoning. Experiments involve backtesting a long-only trading strategy on Chinese A-share stocks from August 2024 to March 2025, comparing FinKario-RAG against vanilla LLMs, financial domain LLMs, and real-world institutional brokerage strategies using metrics like Sharpe ratio, annualized return, and predictive accuracy. Results demonstrate that FinKario-RAG achieves superior performance, outperforming financial LLMs by 18.81% and institutional strategies by 17.85% in predictive accuracy, with a Sharpe ratio improvement of 58.14% over the runner-up. Ablation studies confirm the critical role of the event graph and the two-stage retrieval mechanism, showing significant performance drops when either is removed. The system provides actionable, grounded investment guidance with textual rationales, addressing the lack of explainability in black-box models. Limitations include reliance on the quality of source reports and potential biases in institutional templates, while the scope is currently limited to the Chinese market and specific asset classes.

Research problem

Individual investors lack access to professional-grade analysis and struggle with the volume and complexity of financial information. Existing LLM-based solutions are hindered by two main issues: (1) static knowledge bases cannot keep pace with the rapid evolution of market events, and (2) processing long-form, unstructured equity research reports is inefficient and lacks semantic coherence. The paper aims to automate the construction of a dynamic, event-enhanced financial knowledge graph and develop a retrieval strategy that integrates real-time events with stable fundamentals for accurate stock trend prediction.

Core contributions

• Introduction of FinKario, a dynamic, event-driven financial knowledge graph with over 305,360 entities and 19 relation types, constructed automatically without manual schema definition.
• Development of a dual-schema design (Attribute and Event graphs) guided by professional institutional templates (CFA, FIBO) for structured and interpretable financial knowledge extraction.
• Proposal of FinKario-RAG, a two-stage graph-based retrieval strategy that enhances semantic coherence and context-awareness by retrieving both target entities and their relational neighbors.
• Empirical validation showing FinKario-RAG outperforms financial LLMs and institutional strategies in stock trend prediction accuracy and risk-adjusted returns (Sharpe ratio).
• Creation of a high-quality dataset and pipeline for automated financial knowledge graph construction from unstructured equity research reports.

Method

• Dual-schema KG construction: Separate Attribute Graph (stable fundamentals) and Event Graph (dynamic drivers) built from equity reports.
• Prompt-driven extraction: LLMs guided by CFA handbook and FIBO ontology templates to extract entities and relations automatically.
• Knowledge population: Timestamped extraction of entities and relations from markdown reports, integrated into a unified graph.
• Quality control: Entity normalization, attribute completion via Tushare API, and error correction using LLMs.
• Graph vectorization: Encoding entities, relations, and global graph structure into vectors for semantic search.
• Two-stage retrieval: Coarse-grained retrieval of target entities (stock, date) followed by fine-grained retrieval of related entities/relations.
• Subgraph reconstruction: Mapping retrieved vectors back to the original graph to form a coherent subgraph for reasoning.
• Investment guidance: LLM analyst generates prediction (Rise/Fall), confidence score, and textual rationale based on the subgraph.

Results & findings

• FinKario-RAG achieves a predictive accuracy of 0.581, outperforming the best institutional strategy (Guolian-Minsheng, 0.575) and financial LLMs by 18.81%.
• The method yields an Annualized Rate of Return (ARR) of 2.633, surpassing the runner-up institutional strategy by 30.8% and Stock-Chain by 123.7%.
• Sharpe Ratio reaches 4.926, representing a 58.14% improvement over the best baseline, indicating superior risk-adjusted performance.
• Removing the Event graph causes an 87.2% drop in ARR, highlighting the critical importance of dynamic event data for financial forecasting.
• The two-stage retrieval strategy outperforms Vanilla RAG by 12.70% in predictive accuracy, demonstrating the value of graph-based context expansion.
• FinKario-RAG maintains a moderate Maximum Drawdown (0.172) while achieving high returns, balancing risk and reward effectively.
• Institutional strategies show fragmented performance, with some agencies yielding negative returns, whereas FinKario-RAG provides consistent outperformance.
• The system generates interpretable investment guidance with textual rationales, addressing the explainability gap in black-box LLMs.

Data & experiments

• Dataset: Equity research reports from East Money website (Aug 2024 - Mar 2025), enriched with Tushare financial data and Wind index/industry information.
• Graph Scale: FinKario contains 305,360 entities, 9,625 relational triples, and 19 relation types.
• Backtesting: Long-only trading strategy on Chinese A-share stocks, evaluating weekly signals from Aug 2024 to Mar 2025.
• Baselines: Vanilla LLMs (Qwen3-8B, GPT-4o-mini), Financial LLMs (FinMA, FinGPT, Stock-Chain), and 10 leading Chinese brokerage institutional strategies.
• Metrics: Annualized Rate of Return (ARR), Sharpe Ratio, Maximum Drawdown (MDD), Calmar Ratio, and Predictive Accuracy (ACC).
• Ablation Studies: Analysis of the impact of removing the Event graph, Attribute graph, and using alternative retrieval methods (Vanilla RAG, LightRAG).

Limitations & caveats

• Relies on the quality and bias of source equity research reports, which may reflect institutional biases.
• Scope is limited to Chinese A-share stocks; generalizability to other markets is not tested.
• Performance depends on the accuracy of LLM-based extraction, which may introduce errors in complex cases.
• The backtesting period is relatively short (Aug 2024 - Mar 2025), limiting long-term robustness assessment.
• Does not account for transaction costs or slippage in the backtesting strategy.
• The method requires significant computational resources for graph construction and vectorization.

Curator notes

Discovery status: specific_domain_focus_promoted. Source links: large language models equity research financial analysis