Bacterial foraging trained wavelet neural networks: application to bankruptcy prediction in banks

Abstract

This paper proposes a modified bacterial foraging technique (BFT) to train wavelet neural network (WNN) in order to predict bankruptcy in banks. The BFT is modified in that the swarming step that implements cell-to-cell interaction is deleted. The parameters translation, dilation and the weights connecting different layers in WNN are updated using the BFT. The resulting neural network is called BFTWNN. The effectiveness of BFTWNN is tested on bankruptcy prediction as well as benchmark datasets. We employed ten-fold cross validation in the study. Numerical experiments suggested that the BFTWNN outperformed threshold accepting trained wavelet neural network (TAWNN) (Vinaykumar et al., 2008) and WNN in benchmark datasets by wide margin while it yielded results comparable to that of differential evolution wavelet neural network (DEWNN) (Chauhan et al., 2009) in terms of area under the receiver-operating characteristic curve (AUC). Of particular significance is the superiority of the BFTWNN over the original WNN on all but one datasets.