Research On Financial Early Warning Of Companies Based On Machine Learning And Text Analysis

With the development of financial market,there are more and more the listed companies in China.Because of the fierce competition in external environment,many listed companies face greater financial risks.Changes in the financial condition of listed companies will directly affect the interests of investors and the company's future development.Therefore,it is of great significance to establish an effective financial early warning system for listed companies.Currently,domestic and foreign scholars have made some breakthroughs in the research of financial early warning system.Most of them use machine learning model based on financial data,and have achieved good results.With the development of natural language processing technology,more and more scholars have begun to analyze financial report try to mine information.This paper intends to based on the financial data and financial report text information of listed companies,using accuracy,precision,recall,F1-score,and AUC values as evaluation indicators to establish a financial early warning model.For the early warning model based on financial data,this paper uses the random forest and XGBoost,which are good at processing highdimensional data.We selected financial data as inputs,and used crossvalidation to find the optimal parameters.The final test accuracy of the model was 81%-83%,but the ability in identifying ST samples was not good.For the model based on financial report data,our paper uses Bidirectional Recurrent Neural Network(BRNN)and Self-Attention method.First,we split the words in the chapter of discussion and analysis of operating conditions,then intercepted 250 words as model inputs.We use pre-trained word Embedding,and we get a model accuracy of 80%-82%.Although the accuracy of the model is slightly lower than the financial data model,it has a stronger ST sample recognition ability.Finally,we combine the financial data model with the financial report model by weighted average method.In practical applications,it is acceptable to identify non-ST companies as ST,but identifying ST companies as non-ST may have greater risk,so the recall rate is also an important indicator.In consideration of the comprehensive indicators,we conclude that XGBoost+Self-Attention is the best model.