关键词: 财务舞弊识别, 指标体系, 残差神经网络, 权重分配网络

Abstract: Listed companies are important to capital market. Improving their quality is the basis for ensuring high-quality economic development. Financial fraud will damage the interests of investors, hence, how to accurately identify financial fraud has become an urgent problem to be sovled.
To represent the true financial situation of listed companies, we designed a three-dimensional integrated index system that contains financial indexes, non-financial indexes and quality indexes. Quality indicators can reflect corporate performance and corporate governance, which is an effective supplement to financial indicators and non-financial indicators. To accurately identify financial fraud companies, this study constructs a financial fraud identification model based on residual neural network with a weight distribution network added at the front of the network. This paper selects listed companies in CAMAR database from 1998 to 2022 as data samples. To ensure the stability of the experimental results, the 10-fold cross-validation technique is used to randomly divide the data set for 10 times according to the above proportion, and the average performance of the 10 test results is taken as the final performance of the model. This study adopts the accuracy rate as the performance evaluation index of the model.
To comprehensively analyze the advantages of residual neural network over traditional machine learning methods and ordinary neural network models, two comparison methods are designed in this study, namely “PCA-SVM” and “PCA-BPNN”. Specifically, the popular principal component analysis (PCA) technology is firstly utilized to perform feature transformation on the three-dimensional integrated index system to eliminate the collinearity of the indexes, the obtained principal components are sorted in descending order according to their contribution degree, and then the principal components are selected in turn as the feature components after dimension reduction, until the cumulative contribution degree of the selected feature components is not less than 80%, finally, PCA-SVM and PCA-BPNN are obtained by taking the selected feature components as the input of support vector machine and BP neural network respectively. The experiments show that the performance of the residual neural network model and PCA-BPNN are significantly higher than PCA-SVM (about 12.9%), which shows that the feature transformation and feature modeling capabilities of neural networks are better than traditional machine learning methods. In addition, the accuracy of the model designed in this study is 8.3% higher than that of PCA-BPNN, which indicates that the model based on residual neural network can fully mine and utilize linear or nonlinear information contained in multi-dimensional indexes than ordinary neural network-based model, and has stronger a practical application potential. To analyze the impact of different network structures on the accuracy of final financial fraud identification, this study has designed six models with different network structures. The experimental results show that the network structure used in this study is significantly better than that of the comparison group, and the final accuracy of financial fraud identification reaches 89.3%, showing a strong practical application value.

Key words: financial fraud identification, index system, residual neural network, weight assignment network