Pattern Recognition of Business Failure by Autoassociative Neural Networks in Considering the Missing Values

Abstract

The traditional prediction models of business failure are usually constructed upon the research sample without missing values, that is, the training and testing procedure of the prediction model are not able to be completed if some observations of the relevant variables are missing. This study solves this problem by applying for the data imputation technique of which the autoassociative neural networks and genetic algorithm are consolidated in estimating the missing values. The sample in this study includes 884 Chinese companies listed in Shanghai or Shenzhen stock exchange market during 1996 to 2005, in which sample contains 268 financial distress companies and 616 health companies. There are 38 financial variables and 4 macroeconomics variables used in the model to predict the failure. Sixty percentages of the observations are randomly selected as the training sample, and then the testing sample after randomly deleting 1 to 20 independent variables is further tested. The empirical results show that the average accuracy rate will reach around 78% if number of variables with missing value is controlled by 10 variables. Thus, the proposed AANNGA model is feasible for predicting the business failure in considering the missing values.