Comprehensive U.S. Company Dataset for Financial Statement Fraud Detection
收藏资源简介:
该数据集是由Forvis Mazars与法国学术机构联合构建的综合性美国公司财务欺诈检测基准,整合了结构化财务数据与文本信息。数据集包含10,159条公司季度记录,涵盖2009-2024年期间13,332家公司的财务指标、经大语言模型提炼的管理层讨论与分析摘要文本,并通过人工审核方式关联了美国证监会发布的会计审计执行公告欺诈标签。其构建过程采用自动化提取与专家验证相结合的双阶段流程,确保标签的时序准确性。该数据集专为支撑公司隔离式财务欺诈检测任务而设计,旨在解决传统随机划分方法导致的泛化能力高估问题,推动金融风控领域模型评估范式的革新。
This dataset is a comprehensive U.S. corporate financial fraud detection benchmark jointly constructed by Forvis Mazars and French academic institutions, integrating structured financial data and textual information. It contains 10,159 quarterly corporate records, covering financial metrics of 13,332 firms across the 2009–2024 period, alongside executive discussion and analysis (MD&A) abstracts refined by large language models (LLMs). Fraud labels from the U.S. Securities and Exchange Commission’s (SEC) Accounting and Auditing Enforcement Releases (AAERs) are linked to these records via manual review. Its construction follows a two-stage workflow combining automated data extraction and expert validation, ensuring the temporal accuracy of the fraud labels. Specifically designed to support isolated corporate financial fraud detection tasks, this dataset aims to address the problem of overestimated generalization performance caused by traditional random data partitioning methods, and to drive the innovation of model evaluation paradigms in the financial risk management domain.
数据集概述
该数据集用于金融报表欺诈检测研究,来源于论文《Read Between the Lines: A Robust Financial Statement Fraud Detection Framework》。
数据集内容
- 17,863 份 经提炼的季度管理层讨论与分析(MD&A)报告(使用自托管 QWEN3 32B 模型生成)
- 3,300 份 AAER 报告(会计与审计执法公告)
- 269,097 份 季度财务报告
- 最终可直接使用的预处理数据集,包含三种不同的划分策略:
- 随机划分(Random Splitting)
- 公司隔离划分(Company-isolated Splitting):论文指出该设置是金融报表欺诈检测任务中更严格的框架
- 时间划分(Time Splitting)
每个数据集包含 5 折(folds),每折均提供 train.csv 和 test.csv 文件。
数据集获取
- 预处理数据集:可从 Hugging Face 下载:https://huggingface.co/datasets/WaguyMZ/Financial_statements_fraud_dataset/tree/main
- 完整原始数据集:需通过联系邮箱 guywaffo@gmail.com 获取
数据命名约定
- v4:公司隔离划分的数据集
- v5:随机划分的数据集
- v6:时间划分的数据集




