TLDR: FinKario is a new system that automatically builds and updates a financial knowledge graph from research reports, integrating real-time company data and market events. Combined with its two-stage retrieval strategy (FinKario-RAG), it significantly improves stock trend prediction accuracy, outperforming existing financial AI models and institutional strategies by providing timely and structured financial insights for investors.
In today’s fast-paced financial markets, individual investors often find themselves at a disadvantage, struggling to keep up with the vast amount of information and lacking the professional analysis available to large institutions. Equity research reports are vital resources, offering deep insights into market trends and company performance. However, making the most of these reports with advanced tools like large language models (LLMs) presents two main hurdles: the financial market changes very quickly, and existing knowledge bases can’t update fast enough; and financial reports are often long and unstructured, making it hard for LLMs to integrate information in a timely and relevant way.
To tackle these challenges, a new research paper introduces a groundbreaking solution called FinKario. This system focuses on both the data itself and the methods used to process it. FinKario is a unique dataset that automatically builds a financial knowledge graph, which is essentially a structured network of financial information. It includes over 305,360 entities (like companies or products), 9,625 relational triples (connections between entities), and 19 different types of relationships. What makes FinKario stand out is its ability to automatically pull in real-time company fundamentals and market events. It does this by using a “prompt-driven extraction” method, guided by templates used by financial professionals, ensuring the insights are structured and easy for LLMs to use.
Beyond just creating the dataset, the researchers also propose a smart retrieval strategy called FinKario-RAG (Retrieval-Augmented Generation). This strategy is designed to optimize how LLMs access large and constantly changing financial knowledge, making sure the data retrieved is both efficient and precise. It works in two stages: first, it finds information directly related to a specific query, and then it expands its search to include related entities and their connections, providing a complete financial picture crucial for accurate predictions.
The FinKario system is built with a dual-schema design, featuring an “Attribute Graph” and an “Event Graph.” The Attribute Graph holds stable information about companies, such as their industry, stock exchange, and product lines. This provides a foundational understanding. The Event Graph, on the other hand, captures time-sensitive events like quarterly financial results or key factors driving profitability. This dual structure allows FinKario to dynamically capture both stable financial attributes and evolving market events.
The construction of FinKario involves four main steps: acquiring a domain corpus of financial reports, automatically building the schemas for both the Attribute and Event Graphs using professional templates, populating these graphs with knowledge extracted by LLMs, and finally, a quality control refinement module. This last step ensures the reliability of the extracted information by correcting errors, standardizing entities, and filling in missing data using platforms like Tushare.
Extensive experiments were conducted to test FinKario’s effectiveness, including backtesting its stock trend prediction accuracy against traditional financial LLMs and institutional strategies. The results are impressive: FinKario, when combined with FinKario-RAG, significantly outperforms existing financial LLMs by an average of 18.81% and institutional strategies by 17.85% in stock trend prediction accuracy. Further studies showed that FinKario-RAG itself surpasses other mainstream retrieval methods by an average of 12.70% in predictive accuracy.
The paper highlights several key contributions: the introduction of FinKario as a dynamic, event-driven financial knowledge graph that updates automatically; the proposal of FinKario-RAG, a retrieval strategy that offers a holistic view for financial analysis; and empirical validation demonstrating its superior performance in financial analysis and stock trend forecasting. For more in-depth details, you can read the full research paper here.
Also Read:
- FinWorld: A Comprehensive Open-Source Platform for Financial AI
- Dynamic Knowledge Retrieval: T-GRAG for Time-Sensitive AI Answers
This innovative approach promises to significantly enhance individual investors’ decision-making capabilities and strengthen financial analysis by providing timely, structured, and context-aware financial insights, bridging the gap between rapidly evolving markets and the analytical tools available to everyday investors.
