The prediction of Taiwan government bond yield by neural networks

Numerous studies have demonstrated that the Neural Networks is an accurate and efficient prediction method in various financial applications. Yet research for the field of forecasting government bond yield is still short. Among these limited number of studies, Backpropagation network (BPN) is the most used method. However, suffering form the potential problems, such as slow training speed, long processing time, and possible local minimum, BPN may not be the omni-solution for all applications in practice. The purpose of this research is to provide an in-depth study of effects of on the performance of different neural networks in government bond yields forecasting. Four selected models with different structures, namely Resilient Propagation (RPROP), Radial Basis Function Neural Network (RBFN), Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and Backpropagation network, are investigated and the results are analyzed and compared. The results indicate that (1) the number of nodes in the hidden layer is insensitive to the prediction. (2) The recommended number of input nodes is five. (3) Obviously, more training samples do enhance forecasting performance in our study. (4) The performance of RBFN is the best, followed by ANFIS and RPROP, SVR, and then BPN. (5) BPN is efficient but not the best approach. (6) Our result reveals that RBFN is a useful predicting approach in government bond yield, it performs better than other four models. The recommended combination of parameters for RBFN is five input nodes, six center nodes in the hidden layer, and one output node.