Research On Financial Distress Prediction Driven By Deep Learning

Since the reform and opening up,the development of China's capital market has attracted worldwide attention.With the vigorous development of the capital market,the competition faced by listed companies has become increasingly fierce,and their financial conditions have been severely tested.If an enterprise wants to maintain a stable development in the fierce market competition,it must have a good financial foundation.Once it gets into financial difficulties,it will inevitably affect the development of the enterprise,and even bring the risk of bankruptcy to the enterprise.The bankruptcy of an enterprise will not only endanger the rights and interests of its stakeholders,but also hinder the sustainable growth of the economy,and may even lead to social chaos.In addition,the in-depth development of economic globalization in recent years has brought closer ties between countries and industries,and companies are more likely to fall into financial distress than before and cause greater harm.However,the occurrence of corporate financial distress is usually not an instantaneous event,it is gradual,and often requires a process from financial normal to gradual deterioration,and finally to financial distress.Therefore,if companies can predict whether they will fall into financial distress based on some information in advance,they can take corresponding measures to adjust the company's operations to avoid financial distress.The existing literature mainly uses linear discriminant,logistic regression and machine learning methods to conduct dilemma prediction research.These models have their own limitations.With the development of artificial intelligence and the advent of the era of big data,deep learning methods have gradually been introduced into the field of financial distress prediction.After systematically combing the relevant research literatures of domestic and foreign scholars in the field of financial distress forecasting,this paper finds that the prediction effects of financial distress forecasting models in the previous literature are not uniform and do not take into account the time accumulation of financial distress.Therefore,the cyclic neural network long and short-term memory model(LSTM)in the deep learning method is used to study the financial distress prediction of listed companies,and the interpretability of the model is analyzed.LSTM neural network has the ability to store,read and update long-distance historical information and is good at processing time series data.Theoretically,it has good applicability in the field of financial distress prediction.In terms of sample selection,the sample studied in this paper is A-share listed companies in my country's Shanghai and Shenzhen stock exchanges from 1999 to 2019,and ST companies are regarded as financially distressed companies.In the selection of predictive indicators,referring to the existing literature and combining the principles of comprehensiveness,comparability of data and easy access to data,45 financial and non-financial indicators were finally selected to construct the predictive index system of this article.In the empirical research,first,use data from T-2 years,T-3 to T-2 years,T-4 to T-2 years,and T-5 to T-2 years to construct LSTM neural network financial distress prediction model at different time steps.The prediction results of the LSTM neural network model show that using the sample data from T-3 to T-2 years is the best way to predict whether a listed company will have financial distress in year T.Second,taking into account the imbalance of the sample,the SMOTE oversampling method is used to unbalance the training sample,and the unbalanced data is re-trained and tested for the model.The research shows that the model is unbalanced.The prediction accuracy of positive samples has been significantly improved.The recall rate of the LSTM neural network model from T-3 to T-2 reached 84.15%,and the prediction accuracy of the LSTM neural network model under different time steps has reached more than 90%.Third,using the samples before and after unbalanced to construct a fully connected neural network,random forest and logistic regression financial distress prediction model,and compare the prediction effect of each model with the LSTM neural network.The study found that the prediction effect of the LSTM neural network is better than the other three models,and the imbalance of the sample has different effects on different models.Complex deep learning models are often lack of interpretability and are often used as "black boxes".Therefore,this article finally uses SHAP as a tool to conduct interpretability research on the deep learning model,and obtains the relative importance of the predictive indicators in the model.In addition,the research results show that in the deep learning-based corporate financial distress prediction,financial and non-financial All indicators play an important role.Based on theoretical and empirical analysis,this article believes that the financial distress prediction model based on the LSTM neural network deep learning method has high practical value.