Comparison of Bankruptcy Prediction Models Using Support Vector Machine and Artificial Neural Network

In making financial decisions, many critical factors need to be considered, including the bankruptcy potential of the company. As a result, several bankruptcy prediction models, mainly using machine learning approaches, have been created. Based on relevant research, Support Vector Machine (SVM) and Artificial Neural Network (ANN) often provide the best performance in predicting companies' bankruptcy. On the other hand, both algorithms still have drawbacks, such as SVM is quite sensitive to feature selection, and ANN is prone to overfitting. In addition, many studies ignored the imbalance problems in the bankruptcy dataset, causing the prediction models to be prone to bias. Therefore, this research aimed to compare the performance of the SVM and ANN models in predicting companies' bankruptcy. To overcome the problem of overfitting and feature selection, we utilized the Principle Component Analysis (PCA) dimensionality reduction technique. In addition, for handling the imbalanced dataset problem, we used the Borderline Synthetic Minority Oversampling Technique (BSM). The data used in this study were gained from the annual financial statements of bankrupt and non-bankrupt public companies in Indonesia during the 2010 to 2019 years period. Based on the research, the ANN-BSM-PCA combination model with
$\text{PC =16}$
provided the best accuracy, precision, and recall scores of 0.96, 0.934, and 0.988, respectively. Therefore, the ANN-BSM-PCA with
$\text{PC =16}$
was the best model in the case of predicting companies' bankruptcy.