基于集成学习和文本分析的财务欺诈识别研究

Research on Financial Fraud Recognition Based on Integrated Learning and Text Analysis

管理层讨论与分析(MD&A)作为公司年报的重要组成部分,在财务欺诈识别中具有不可或缺作用。选取2019—2020年沪深两市A股制造业中具有财务欺诈行为的上市公司作为研究样本,对年报中MD&A所披露的信息进行度量,实现中文文本信息的量化;并在对结构化文本数据和定量数据进行融合的基础上提出基于Stacking集成算法的财务欺诈识别模型。实验结果表明:Stacking集成分类器在准确率、查准率、F1-score以及AUC这4个评价指标得分上均取得了最优值,识别性能显著优于传统单分类器;相较于使用传统定量数据,MD&A中文文本信息的加入使Stacking集成分类器的识别性能得到显著提升。应充分挖掘年报文本信息,并持续优化模型和算法,以提升财务欺诈识别系统的准确性。

As an important part of a company’s annual report,Management Discussion and Analysis(MD&A)plays an indispensable role in the identification of financial fraud.By selecting listed companies with financial fraud in Shanghai and Shenzhen A-share manufacturing industry in 2019-2020 as research samples,the information disclosed in MD&A in annual reports is measured to realize the quantification of Chinese text information;Furthermore,a financial fraud recognition model based on the integration algorithm of structured text data and quantitative data is proposed based on the integration of structured text data and quantitative data.The experiment shows the following results.The integrated classifier has optimal values in accuracy,precision,F1-score,and AUC,and its recognition performance is significantly better than that of traditional single classifiers.Compared with traditional quantitative data,the addition of MD&A Chinese text information significantly improves the recognition performance of the integrated classifier.The text information of the annual report should be fully mined,and the model and algorithm should be continuously optimized to improve the accuracy of the financial fraud identification system.