A hybrid bankruptcy prediction model based on GMDH-type neural network and genetic algorithm for Tehran Stock Exchange companies

This paper proposes a Soft Computing model for effective bankruptcy prediction, based on the integration of Group Method of Data Handling (GMDH) neural network and genetic algorithm which is called here as GA-GMDH. Genetic algorithm (GA) designs the whole architecture of the GMDH network and optimizes its topology. In order to demonstrate the effectiveness of our proposed GA-GMDH model, its performance was compared with performance of the commonly used statistical techniques of logistic regression (LR) and a relatively new artificial intelligent technique of Adaptive Neuro-Fuzzy Inference System (ANFIS). Performance of the designed prediction models depends on the utilized variable selection technique. Therefore, we constructed 12 prediction models through combining the four filtering feature selection methods and the three prediction models. The four feature selection methods of independent samples T-test, correlation matrix (CM), stepwise multiple discriminant analysis (SDA) and principal component analysis (PCA)are combined with prediction models to generate four optimal feature subsets. Empirical data were collected one year prior to failure from Tehran Stock Exchange (TSE) during 1997-2008. For robust assessing of prediction models’ performance, we applied Type-I and Type-II errors, and area under the receiver operative characteristics curve (AUC) measures. Experimental results indicate that our proposed GA-GMDH model has high ability in bankruptcy prediction problem and significantly outperforms ANFIS and LR models in all combinations with four feature selection methods. Meanwhile, the CM method has the best ability in selecting predictive variables in comparison with other feature selection methods. Therefore, CM-GA-GMDH model is determined as the best constructed model for bankruptcy prediction using our particular dataset from TSE.