New Tuning Methods for the Architecture of Neural Network Model with Application to Bankruptcy Prediction

• Activity Number: 637
• Type: Contributed
• Date/Time: Thursday, August 5, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Chulwoo Jeong*+ and Myung Suk Kim and Jae H. Min
• Companies: Sogang University and Sogang University and
• Keywords: bankruptcy prediction ; neural network model ; generalized additive model ; genetic algorithm ; grid search method ; tuning parameters

Abstract

The performance of a neural network model is known to be affected by its important elements such as input variables, number of hidden nodes, and value of decay constant. In this paper new approaches to tune these factors are suggested for more improved accuracy. For the input variable selection, the method via the generalized additive model (GAM) is applied. The grid search method and the genetic algorithm are implemented to tune the number of hidden nodes and the value of weight decay parameter. The neural network models tuned by the suggested methods are applied to the bankruptcy prediction, and are compared with existing popular bankruptcy forecasting models such as the GAM, the generalized linear model, and the support vector machine. Our empirical results indicate that the newly tuned neural network model seems to significantly outperform the aforementioned models.