Research On Financial Crisis Early-warning Of Listed Companies Based On Features Imaging And Convolutional Neural Networks

Recently,with the intensification of Sino-US trade negotiations and the continuous fluctuation of the market environment,enterprises often encounter various problems such as financial deterioration in actual management.The emergence of financial problems is often fatal to the development of the entire company,which makes the financial crisis gradually become the focus of attention in economies closely related to the development of enterprises.Therefore,the establishment of an effective and scientific financial early warning model is a research direction of risk management in the process of enterprise operation management.However,limited by data analysis and processing technology,traditional financial early warning is mostly based on simple statistical Z-score model analysis,support vector machine and BP neural network model.In the process of model construction and use,there will be incomplete selection of sample data,scientific selection of index variables,"unification" of the modeling process and other issues.With the advent of the era of big data,deep learning-related technologies can be used to achieve "unification" of financial early warning modeling,reduce human intervention,and establish more scientific financial early warning models,which has important research significance.Based on a review of domestic and foreign scholars’ research on financial early warning research,this article selects the annual reports from the Wande database in accordance with five aspects reflecting the solvency,operating ability,profitability,corporate growth ability and cash flow ability of listed companies.The characteristics of financial indicators,and at the same time selected other financial analysis data in the database that can reflect the financial situation changes with actual financial significance,such as total asset turnover,equity multiplier,etc.,and constructed anoriginal indicator system including 194 indicators.Then select the T-2 and T-3financial statement data of 4,210 listed companies as the source data.For the financial early warning model of listed companies based on convolutional neural network,this paper uses MATLAB2018 to convert the original 194 input index variables into a 14 *14 feature image,and then optimizes the classic LeNet-5 convolutional neural network as needed,thereby Construct a financial early warning model based on convolutional neural network and train and evaluate the model.Subsequently,the original 194 index variables were subjected to statistical tests such as difference and collinearity to eliminate the "inappropriate" part of the index variables to achieve the screening of the index variables.The traditional classic support vector machine and BP neural network financial early warning model were used for comparison model.The experimental results show that the G-mean and AUC values of the financial early warning model based on convolutional neural network reach 78.6% and 92.8%,which are different from traditional support vector machine and BP neural network-based early warning models.Through the research,compared with the traditional early warning model based on support vector machine and BP neural network,the financial early warning model proposed in this paper based on feature mapping into grayscale maps using CNNs can not only effectively avoid the influence of subjective factors on the modeling,And has more accurate early warning evaluation performance,which further improves the stability and applicability of the model.