The Role of Visual Financial Ratios in Predicting Corporate Bankruptcy Using Convolutional Neural Network Models and Comparing them with Traditional Models

The purpose of this study is to test the use of visual financial ratios to predict the bankruptcy of companies using a convolutional neural network and compare it with traditional models.

The research period was 2009 to 2018. The sample companies have been selected from the ones which were listed on the Tehran Stock Exchange in two groups of bankrupt companies (66) and non-bankrupt companies (66). Since the work of convolution neural network is to recognize images from existing images, first the financial ratios were converted into images as research data through MATLAB 2019 software, then, the situation of the sample companies were predicted and diagnosed with the help of convolution neural network and under Google net architecture.

Convolutional neural network models performed accurate images and predictions with 50% accuracy. On the one hand, in order to strengthen the results and determine the effectiveness of the first hypothesis, three other hypotheses were proposed to be compared to Altman, Spring-gate and Zimski models. The results of all three indicated that the convolution model was not confirmed as accurate compared to these three models.

Advances in computers and the use of deep learning, which is a kind of improvement in artificial intelligence, affect the prediction of bankruptcy through visual financial ratios. However, to consolidate the test results of the first hypothesis, three practical models of bankruptcy prediction including Altman (1983), Springgate (1978) and Zimski (1984) were tested, the results of which did not confirm the accuracy of the convolution model compared to these three models.