Failure Prediction for South African Companies

The aim of this paper is to construct and investigate a superior model that can predict bankruptcy for JSE listed companies using a traditional statistical method (Logistic Regression, LR) and artificial intelligent related models (Support Vector Machines, SVM; Random Forest, RF). This research adopted the paired sample approach for 144 companies. 72 companies that are delisted from JSE due to business failure or scheme of arrangement from 2005 to 2018 are matched with 72 bankrupt counterparts based on year of delisting, industry and asset size. Our results indicate that compared with other models investigated, SVM-RBF outperformed other models for both the training dataset and testing dataset with average accuracies of 83.3% and 65.8% respectively for the t-test selected input variables. For the two-step feature selection (stepwise logistic regression) input variables, SVM-RBF has the highest average accuracy performance of 71% for the training data. For the testing data, SVM-RBF and SVM-linear show the highest average performance of 65%. This paper finds evidence in support of artificial intelligence related techniques such as SVM models having better performance compared with other models examined in this paper. This model can be employed as a tool for predicting failure five years before hand by investors and academics may refer to it as a fundamental reference material.