Research On Enterprise Financial Distress Prediction Method Based On Deep Learning

In the field of enterprise risk management,it has been an important research topic to predict whether an enterprise is in financial distress.When an enterprise is in financial distress,it may default on its debts and be in a state of loss for several consecutive years.More serious financial difficulties will lead to enterprises applying for bankruptcy,and even affect the entire industry and financial market.An effective financial distress prediction model can give an early warning before the occurrence of financial distress,which is very necessary in today’s highly uncertain environment.The prediction methods of financial distress have developed from the early expert discrimination and classic statistical methods to machine learning and deep learning models.Data selected have also changed from static data to time series data.However,at present,most of the research on financial distress prediction only process the time series data at a single time scale.How to mine more abundant time series information is still lack of research and exploration.Additionally,the relevant data of enterprises is gradually diversified and more available.How to effectively integrate the information in time series with different length is the key issue to be solved in the financial distress prediction task.In addition,machine learning models are far from human intelligence in terms of advanced cognition and complex decision-making.How to improve the applicability and reliability of machine learning models in assisting human decisionmaking analysis is an important issue faced by enterprise financial distress prediction scenarios.In view of the above problems,domestic listed companies are selected as the research object and the paper carries out the following research work:Firstly,this paper designs two multi-scale information representation models.For the multi-scale information representation model of financial index data,this paper decomposes the hidden state of a single-layer LSTM into multiple parts,and each part is responsible for extracting a feature representation of a time scale.At the same time,the model adaptively learns the scale weights at different time steps through the Cross-Attention mechanism.For the multi-scale information representation model of stock index data,this paper improves the Multi-Head Attention mechanism in Transformer to the Multi-Scale Multi-Head Attention mechanism,so that each head focuses on a time scale learning.This paper uses financial index data and stock index data to test the two models designed.Compared with other models,the macro-F1 score of this model is improved.Secondly,this paper designs a multi-source asynchronous time series data fusion model,mainly including a multi-source information interaction module and a gate structure fusion module.The multi-source information interaction module learns the information association between two time series through Cross-Attention mechanism to realize the interaction of information.Then,a gate structure is used to control the flow state of the two sources of information and realize the fusion of the two sources of time series data from the feature level.The experimental results demonstrate that the macro-F1 score of the model is better than other data fusion methods used in the field of financial distress prediction,which reaches 0.933.Finally,this paper proposes a framework for predicting the financial distress of enterprises based on the organizational "Human-in-the-loop"paradigm.In this framework,humans play the role of organizers.This framework allows human intelligence and machine intelligence to work together,and improves human participation and involvement in the whole process.At the same time,some new evaluation indexes are proposed for this framework from the perspective of human-machine cooperation adaptability.The effectiveness and feasibility of the designed framework is verified by recruiting volunteers to conduct behavioral experiments.As a whole,an effective attempt is made for the further development and research of the "Human-in-the-loop" paradigm in the field of enterprise financial distress prediction.