Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FinStat2SQL: A Text2SQL Pipeline for Financial Statement Analysis

Quang Hung Nguyen, Phuong Anh Trinh, Phan Quoc Hung Mai, Tuan Phong Trinh

TL;DR

The paper presents FinStat2SQL, a lightweight, domain-specific text2sql pipeline tailored to Vietnamese financial standards (VAS) that combines a multi-agent NL-to-SQL flow with a domain-focused database and synthetic QA data. It demonstrates that fine-tuned, smaller models can rival larger proprietary systems when paired with vector-based retrieval, self-correction, and decomposition strategies, achieving strong accuracy and sub-4-second latency for practical use in Vietnamese enterprises. The work also introduces a hybrid, finance-aware evaluation framework and a robust dataset, showing that a 7B finetuned model can perform competitively, while proprietary Gemini models achieve the highest accuracy (around 72%), highlighting the trade-offs between cost and performance. While offering a practical solution for local financial analytics, the study notes limitations in scope and terminology and points to future work on broader coverage, cross-national standards, and predictive analytics to extend applicability and impact.

Abstract

Despite the advancements of large language models, text2sql still faces many challenges, particularly with complex and domain-specific queries. In finance, database designs and financial reporting layouts vary widely between financial entities and countries, making text2sql even more challenging. We present FinStat2SQL, a lightweight text2sql pipeline enabling natural language queries over financial statements. Tailored to local standards like VAS, it combines large and small language models in a multi-agent setup for entity extraction, SQL generation, and self-correction. We build a domain-specific database and evaluate models on a synthetic QA dataset. A fine-tuned 7B model achieves 61.33\% accuracy with sub-4-second response times on consumer hardware, outperforming GPT-4o-mini. FinStat2SQL offers a scalable, cost-efficient solution for financial analysis, making AI-powered querying accessible to Vietnamese enterprises.

FinStat2SQL: A Text2SQL Pipeline for Financial Statement Analysis

TL;DR

The paper presents FinStat2SQL, a lightweight, domain-specific text2sql pipeline tailored to Vietnamese financial standards (VAS) that combines a multi-agent NL-to-SQL flow with a domain-focused database and synthetic QA data. It demonstrates that fine-tuned, smaller models can rival larger proprietary systems when paired with vector-based retrieval, self-correction, and decomposition strategies, achieving strong accuracy and sub-4-second latency for practical use in Vietnamese enterprises. The work also introduces a hybrid, finance-aware evaluation framework and a robust dataset, showing that a 7B finetuned model can perform competitively, while proprietary Gemini models achieve the highest accuracy (around 72%), highlighting the trade-offs between cost and performance. While offering a practical solution for local financial analytics, the study notes limitations in scope and terminology and points to future work on broader coverage, cross-national standards, and predictive analytics to extend applicability and impact.

Abstract

Despite the advancements of large language models, text2sql still faces many challenges, particularly with complex and domain-specific queries. In finance, database designs and financial reporting layouts vary widely between financial entities and countries, making text2sql even more challenging. We present FinStat2SQL, a lightweight text2sql pipeline enabling natural language queries over financial statements. Tailored to local standards like VAS, it combines large and small language models in a multi-agent setup for entity extraction, SQL generation, and self-correction. We build a domain-specific database and evaluate models on a synthetic QA dataset. A fine-tuned 7B model achieves 61.33\% accuracy with sub-4-second response times on consumer hardware, outperforming GPT-4o-mini. FinStat2SQL offers a scalable, cost-efficient solution for financial analysis, making AI-powered querying accessible to Vietnamese enterprises.

Paper Structure

This paper contains 24 sections, 5 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Problem in Financial Standards convergence
  4. Text2SQL
  5. Preliminaries
  6. Database Construction
  7. Synthetic QA Dataset
  8. Evaluating Dataset
  9. Approach
  10. Text2SQL Pipeline
  11. Metrics
  12. Experiment
  13. Setup
  14. Result
  15. Conclusion
  16. ...and 9 more sections

Figures (5)

  • Figure 1: Diagram of the FinStat2SQL Pipeline: The language model first parses and analyzes the user's prompt, retrieves relevant company database information, and then iteratively executes and debugs SQL queries until a correct result is obtained, which is returned as the final answer. More information can be found in Sec. \ref{['sec:Approach']}.
  • Figure 2: Diagram of Entity Extraction and Row Selection process
  • Figure 3: SQL generation process
  • Figure 4: Question Answering with FinStat2SQL
  • Figure :