Neural network-based systems that allow the system, through an analysis of historical data, to determine the relationship between account characteristics and the probability of default. The backpropagation algorithm - the multilayer feedforward network structure is described based on data with nine financial ratios from 81 firms listed in China. A simulation on network is made. Neural network model for classification algorithms are established. By varying network parameters we demonstrate that LevenbergMarque training error is smallest among 4 learning algorithms and its performance is better. Increasing the number of hidden layer can result in minor improvement