The effect of training set distributions for supervised learning artificial neural networks on classification accuracy

Neural networks have been repeatedly shown to outperform traditional statistical modeling techniques for both discriminant analysis and forecasting. While questions regarding the effects of architecture, input variable selection, learning algorithm, and size of training sets on the neural network model's performance have been addressed, very little attention has been focused on distribution effects of training and out-of-sample populations on neural network performance. This article examines the effect of changing the population distribution within training sets for estimated distribution density functions, in particular for a credit risk assessment problem.