Agentic and Event Enhanced Financial Knowledge Graph Approaches

Introducing FinReflectKG

While FinKario demonstrates event integration, another line of work aims to generalize and evaluate knowledge graph construction from corporate filings. The FinReflectKG project (October 2025) builds a large scale financial knowledge graph from SEC 10‑K filings of the S&P 100 companies. The authors note that most financial knowledge graphs rely on news feeds and lack rigorous evaluation. To address this, they propose an agentic extraction framework combining four components: intelligent document parsing, table aware semantic chunking, schema guided iterative extraction and a reflection driven feedback loop. This system supports three extraction modes, single pass, multi pass and reflection agentic, allowing trade offs between efficiency and accuracy. Empirical evaluation shows that the reflection agentic mode attains a 64.8 % compliance score across rule‑based policies and outperforms baseline methods on precision, comprehensiveness and relevance.

In addition to the extraction pipeline, FinReflectKG provides a holistic evaluation framework that uses rule‑based checks, semantic diversity metrics and a “LLM‑as‑a‑Judge” approach to benchmark the quality of extracted triples. By releasing the dataset and evaluation framework, FinReflectKG promotes transparency and reproducibility.

FinReflectKG‑MultiHop benchmark

A companion benchmark, FinReflectKG, MultiHop, focuses on multi‑hop question answering over financial disclosures. Questions often require connecting facts across different sections, filings and years, which is challenging for LLMs operating on raw text. The MultiHop benchmark links audited triples from the FinReflectKG dataset to their source text spans and generates analyst style questions by mining frequent 2 and 3 hop subgraphs. During evaluation, the authors compare three retrieval strategies: (S1) exact knowledge graph linked paths, (S2) text only page windows and (S3) page windows with random distractors. Precise knowledge graph guided retrieval increases correctness by ≈24% and reduces token usage by ≈84.5% compared with the vector retrieval baseline.

Comparing FinKario and FinReflectKG

Both FinKario and FinReflectKG automate financial knowledge graph construction but target different data sources and use cases. FinKario extracts event centric triples from research reports and integrates them with company fundamentals. FinReflectKG, on the other hand, processes regulatory filings (10‑K reports) and emphasises evaluation and compliance. FinKario introduces a retrieval augmented generation component (FinKario‑RAG) to deliver relevant subgraphs to LLMs, whereas FinReflectKG focuses on schema guided iterative extraction and reflection driven feedback.

Summary

1. FinReflectKG uses intelligent parsing, table‑aware chunking, schema‑guided extraction and reflection‑driven feedback to build a comprehensive KG from SEC filings.

2. FinReflectKG, MultiHop demonstrates that KG‑guided retrieval dramatically improves multi‑hop question answering accuracy and efficiency.

3. FinKario and FinReflectKG tackle different sources (research reports vs regulatory filings) but together highlight the trend toward automated, event‑aware and evaluable financial KGs.