A Dynamic Financial Distress Forecast Model Under Unbalanced Data Environment

Corporate financial distress forecast is an important research topic in the field of corporate credit evaluation,and it is also a thorny problem faced by corporate management,investment,credit granting and regulatory decision-making in real economy.Untrustworthy behaviors often occur in financial industry and social field.Therefore,assessing and predicting the credit risk of an enterprise is of great significance to companies,investors and regulatory authorities.Based on a systematic review of theoretical research,model research,and model evolution of financial distress forecasting by domestic and foreign scholars,this paper grasps the technical development context of financial distress forecasting,defines the connotation of financial distress,summarizes the performance and causes of financial distress and generalizes the main problems and deficiencies existing in the field of current financial distress forecasting.And found that,as most financial distress forecast(FDF)models in previous studies tended to ignore the two key financial distress data characteristics,imbalanced datasets and data stream characteristics.The imbalanced datasets cause the model built on balanced data to have low classification accuracy on unbalanced samples;the characteristics of the data stream and the resulting concept drift cause the accuracy of the old model to decrease over time on the new data set.To overcome the above two problems,this paper proposes a new dynamic financial distress forecasting algorithm(ANS-REA algorithm),which deals with the imbalance problem at the sample cluster level.And the imbalance processing method combining historical sample addition and oversampling solves the absolute imbalance problem of the minority clusters.At the same time,in order to solve the problem of concept drift,this paper uses a sliding time window mechanism in sample selection and uses predicted AUC values as weights to integrate subclassifiers established on different time stamp samples.In addition,we also pays attention to the staged characteristics of financial distress,and proposes a new dynamic financial distress forecasting framework that combines multi-period forecast results.The framework selects forecast indicators in the T-2 to T-6 periods,and uses an index weighting function to weight and integrate the forecast results in each period to form the final forecast result.In terms of empirical experiment: This paper selects 373 listed companies that were ST classified because of financial anomalies from 2007 to 2017 as a sample of financial distress.And using the propensity matching score method(PSM method)to match the financial distress sample with 1119 normal financial companies.The two samples together constitute the original prediction sample with an imbalance ratio of 1: 3.In the construction of forecast indicators,we selected 53 candidate financial indicators and 17 candidate non-financial indicators.The financial indicators cover the company's long-term and short-term solvency,development ability,operating ability,profitability and cash flow.Non-financial indicators include corporate governance structure,equity concentration,managment holdings and state holdings.Then,we select the forecast indicators of the T-2 to T-6 periods by significance test.The empirical results show that the application of the ANS-REA method can effectively increase the number of minority samples,alleviate the imbalance of samples,and not blur the boundaries between the minority samples and the majority samples.From the comparison of the performance of different basic classifiers,the random forest algorithm(RF)is superior to the other six commonly used base classifiers.From the comparison of the effects of different imbalanced processing methods,the ANS-REA algorithm has the best performance.From the comparison of the effects of forecasting frameworks combined with different periods of forecasting results,the financial distress forecasting framework combined with the five-period forecast results(T-2 to T-6)achieved the highest predicted average AUC value.Finally,we put forward five suggestions based on the research conclusions.