A tuning method for the architecture of neural network models incorporating GAM and GA as applied to bankruptcy prediction

The performance of a neural network model is affected by important constituent elements such as input variables, the number of hidden nodes, and the value of the decay constant. This paper suggests a new approach to fine-tune these factors to improve their accuracy. For the input variable selection, the generalized additive model (GAM) is applied. The grid search method and the genetic algorithm are sequentially implemented to fine-tune the number of hidden nodes and the value of the weight decay parameters. This suggested method to improve the neural network model is used to predict the probability that a firm may apply for bankruptcy, and its performance is compared with the results of existing bankruptcy forecasting models such as case-based reasoning, the decision tree, the GAM, the generalized linear model, the multi-variate discriminant analysis, and the support vector machine. Our empirical results indicate that the newly tuned neural network model significantly outperforms the other models.