Comparison And Improvement Of Models Evaluating Management Fraud Risk

Since the securities market was born, a lot of management fraud cases in listed companieshave happened, without the trend of disappearance. This severely breaks the order of the securitiesmarket and causes great damage to the investors and society. Finding an effective and efficient wayto identify management fraud and prevent it happening is so urgent and meaningful. Therefore,under such a background, this thesis compares and analyses several fraud risk assessment modelsand puts forward a new idea to evaluate the risk of management fraud.Above all, this thesis summarizes and analyses the existing research achievements in the fieldof management fraud to look for significant fraud identification signals and effective fraud riskassessment models. Foreign and domestic literatures review contains four aspects: definition ofmanagement fraud, fraud motivation theories, identification signals of management fraud, andidentification ways of management fraud. Secondly,255Chinese listed companies manipulatingmanagement fraud from2008to2012and255non-fraud companies are chosen as research samples.Thirdly, this thesis builds a comprehensive and systematic index system of management fraud riskevaluation as the input variables of risk assessment models. The variables in the index system arefrom six perspectives such as ownership structure, corporate governance, particular transactions andevents, operating pressure, the relationship with auditors, and finance stability. Then the510samples mentioned before are divided into two groups, one group containing370companies iscalled training group and the other is called testing samples. All the training and testing samples areput into four risk assessment models: Logistic regression model, BP neural network model, LVQneural network model and SVM model. After comparing and analyzing the accuracy, advantagesand disadvantages of the four risk assessment models, this thesis puts forward a new risk evaluationidea of building a combined classifier. Finally the thesis makes a conclusion based on thetheoretical analysis and empirical research.The contributions of this thesis are significant. First of all, the thesis establishes an indexsystem including financial and non-financial variables to compare the accuracy of different riskassessment models. Another contribution is that this thesis comes up with a new model-acomprehensive model based on classifier combination to evaluate the management fraud risk. Thecomprehensive model with its accuracy of78.6%can not only improve the overall accuracy butalso increase reliability of the assessment result. It has great value of application.