管理工程学报2026，Vol.40Issue(5)：61-76,16.DOI:10.13587/j.cnki.jieem.2026.05.005
企业财务违规的机器学习识别模型
A machine learning identification model for corporate financial misconduct
摘要
Abstract
The presence of corporate financial misconduct can greatly disrupt the smooth functioning and fairness of the capital market,in turn affecting the stability of the broader economy.Due to the growing stealthiness of such malpractice,its occurrence has been on the rise.In the past,researchers have relied on empirical analysis,univariate tests,and regression analysis to uncover the factors driving corporate financial wrongdoing.However,with technological progress,the use of innovative techniques like machine learning to detect financial misconduct has become a recent trend.Though existing studies mainly center on financial fraud and tend to rely on a single methodology to scrutinize a particular aspect,such an approach is insufficient to accurately identify risks of financial misconduct.Consequently,it fails to prevent the potential negative consequences of these occurrences right from the start.Hence,it is imperative to develop a framework that can effectively identify a wider array of corporate financial misconduct using publicly accessible data.This is essential for protecting the interests of various stakeholders,including regulators,investors,creditors,and upstream and downstream supply chain companies. This paper attempts to construct a model to accurately identify corporate financial misconduct risks by applying it to a typical case to demonstrate its application process clearly and intuitively,so as to provide an effective analytical tool for building a fair and efficient capital market.Specifically,this paper first briefly reviews the financial misconduct case of Kaile Technology.Under the guidance of the fraud triangle theory,a preliminary selection of indicator items is made from the three dimensions of opportunity,excuse,and pressure.After using the non-compliant listed companies and their matched samples to eliminate insignificant items,a refined index system is developed using algorithmic processes,ensuring its theoretical robustness,scientific integrity,reliability,and practicality.Subsequently,machine learning methods such as the Naive Bayes model,CART decision tree model,gradient boosting machine model,and XGBoost model are used to estimate parameters separately,and the most suitable combination for each scenario is selected after comparing various estimation results.Additionally,the constructed financial misconduct identification model is applied to the case of Kaile Technology to demonstrate the working process of the model.Finally,the model is used to estimate a randomly selected sample of listed companies to verify its universal applicability.The research findings indicate that the identification model constructed in this paper possesses a high level of credibility and validity and is capable of accurately identifying the financial misconduct risks of companies like Kaile Technology,and of elucidating the underlying causes of such incidents from the three dimensions of opportunity,excuse,and pressure. This paper makes incremental contributions to the existing research in terms of research paradigms,the constructed financial misconduct identification model,and the perspectives adopted.First,it adopts a research paradigm that combines large-sample archival data with typical case studies,innovatively applying a large-sample empirical method to construct a financial misconduct identification model and then applying the model to the typical case of Kaile Technology.This approach intuitively demonstrates the working process of the constructed model,meeting the needs of various stakeholders to understand the application of emerging methods such as machine learning.Second,it uses the well-established fraud triangle theory as the preliminary selection standard for indicator items,ensuring high theoretical relevance while significantly reducing any subjective biases.Moreover,it innovatively introduces unstructured data and external environmental factors into the indicator system,broadening the sources of indicator items for financial misconduct identification.At the same time,after comparing various machine learning identification methods,it selects the most suitable estimation method for each scenario and forms a combination of estimation methods,ensuring that the constructed model is both scientifically sound and feasible.Third,the paper constructs an identification model from a broader perspective,focusing on the increasingly frequent and harmful corporate financial misconduct in the capital market,thereby expanding the model's application scope. Overall,against the backdrop of the severe and frequent occurrence of corporate financial misconduct in China's capital market,this paper constructs a robust model that is not only academically rigorous and scientifically reliable,but also practical for real-world applications for identifying financial misconduct risks.The model is applied in a typical case study to open up the"black box"of its workings,providing analytical tools and insights to foster the stable and efficient growth of the capital market.关键词
财务违规/机器学习/识别模型/三角形理论Key words
Financial misconduct/Machine learning/Identification model/Fraud triangle theory分类
管理科学引用本文复制引用
黄英,康晓佳,唐凯..企业财务违规的机器学习识别模型[J].管理工程学报,2026,40(5):61-76,16.基金项目
浙江大学-悉尼大学种子基金项目(188170*194252404/002) （188170*194252404/002）
中国统一战线理论研究会浙江研究基地 USyd-ZJU Partnership Collaboration Awards(188170*194252404/002) （188170*194252404/002）
The Zhejiang Research Base of China United Front Theory Research Association （）
