Configuring an improved backpropagation network for forecasting study of interest rate in traditional money market and derivative commodity market

Abstract

By good management of the interacting characteristics of micro and meso structures for optimizing performance of feedforward networks, the application of a neural network to pattern recognition of monetary tools, bond rating, stock price forecasting and loan examination has successfully been done. The study focuses on the prediction of future trends of the 90 to 180 day commercial paper interest rate. The outcome shows several encouraging messages: (1) While the result of applying the multiregressional model on this kind of problem is awkward, the improved backpropagation networks, especially the one integrating Nguyen-Widrow Method and Adaptive Learning Rate Method have good performance without involving the serious problems of multicollinearity and autocorrelation. (2) With small tolerance error, the network forecasting reliability is satisfactory no matter whether random or moving simulation sampling is adopted. (3) For avoiding the impact of random wave, we take the average daily interest rate t-2,t-1,t+1,t+2 as the target output. In so doing the network presents a good learning effect with the accuracy of forecast beyond 98%. (4) The performance of the improved backpropagation network like momentum is not always better than a pure backpropagation network. We learned from the study that the fluctuating trend of interest rate may be influenced by different combinations of economic and monetary independent variables in different time periods, so rashly gathering a big sample without reviewing the attributes of the data may prevent the authentic forecasting effect of the network.