The primary goal of this paper is to get over the limitations of single neural network models through model integration so as to increase the accuracy of bankruptcy prediction. We take the closeness of the output value to either 0 or 1 as the model's confidence in its prediction as to whether or not a company is going to bankrupt. In case where multiple models yield conflicting prediction results, our integrated model takes the output value of the highest confidence as the final output. The output of the confidence-based integration approach significantly increases the prediction performance. The results of composite prediction suggest that the proposed approach will offer improved performance in business classification problems by integrating case-specific knowledge with the confidence information and general knowledge with the multi-layer perceptron's generalization capability.