FinReporting: Agentic Localized Reporting Workflow for Cross-Jurisdictional Financial Disclosure

FinReporting is an agentic workflow system for cross-jurisdictional financial reporting, designed to address the semantic alignment challenges caused by differences in accounting standards, tagging infrastructure, and aggregation practices across markets. By constructing a unified normative ontology and an auditable multi-stage processing flow, its consistency and reliability have been verified using annual report data from the U.S., Japan, and China, providing global investors with a transparent and auditable cross-market financial analysis infrastructure.

Against the backdrop of globalized investment, cross-market financial reporting faces three structural differences: 1. Diversity of accounting standards: Fundamental differences exist in revenue recognition and other areas among U.S. GAAP, Japanese JGAAP, and Chinese CAS; 2. Fragmentation of tagging infrastructure: XBRL structured tagging (U.S. and Europe), PDF native disclosure (Japan), and hybrid models (China) lead to varying difficulties in information extraction; 3. Implicit differences in aggregation practices: The same transaction may be categorized differently in reports across markets (e.g., expense item classification). These differences easily lead to semantic misalignment and verification difficulties when LLMs process data automatically.

The core innovations of FinReporting lie in its unified normative ontology (a semantic alignment framework covering the three core financial statements) and four-stage auditable workflow: 1. Document acquisition: Automatically obtain original documents from exchanges/official websites and record logs; 2. Information extraction: Adopt differentiated strategies for XBRL/PDF/HTML; 3. Normative mapping: Map raw data to the unified ontology based on understanding of accounting standards; 4. Exception logging: Record items that cannot be automatically aligned for manual review. Additionally, the system uses LLMs as constraint validators, requiring their decisions to be supported by accounting standard clauses.

The research team verified the system using annual report data from listed companies in the U.S., Japan, and China: In terms of consistency, the account alignment accuracy increased by approximately 35% compared to keyword matching, and the numerical extraction error rate decreased by about 60%; In terms of reliability, the anomaly detection coverage reached 90%, only 5% of items required manual intervention, and the average processing time per annual report was within 30 seconds. The test covered different industry sizes to ensure the generalizability of the results.

Application scenarios include global equity funds (tracking multi-market companies), cross-border M&A analysis (quickly understanding target finances), RegTech (regulatory cross-border compliance), and academic research (large-scale cross-country comparisons). The team has released an interactive demonstration system that supports selecting markets/companies, viewing standardized reports, downloading structured data, and tracing original sources to ensure transparency.

Limitations: Limited coverage of emerging markets (e.g., India, Brazil), difficulty in processing non-financial information (ESG/management discussion), and need for optimization in real-time performance (quarterly reports/ad-hoc announcements). Future directions: Dynamic ontology evolution (adapting to accounting standard updates), multilingual enhancement (improving Japanese and Chinese text understanding), and predictive analysis (financial health scoring/risk early warning).

FinReporting represents a paradigm shift in AI applications in the financial sector—from black-box intelligence to auditable intelligence. Its ability to cross jurisdictional boundaries, understand different accounting standards, and maintain transparency will become a key infrastructure for institutional investors. For practitioners and researchers, it demonstrates an example of combining LLMs with rigorous engineering practices to solve complex problems.