Integrating multivariate statistics and neural networks for interest rate prediction

Abstract

Statistical techniques and software learning methods have been used extensively to understand financial market behavior. Numerous studies reveal that each of these techniques can provide good results. However, an integrated methodology combining several multivariate techniques may offer even better performance. To explore this conjecture, a variety of integrated strategies and predictive methods were considered in forecasting 10-year Treasury notes. The preprocessing methods were as follows: straightforward standardization of input data, reduction of input variables by stepwise regression and factor analysis, selection of input variables`` lags by diagonal correlation analysis. The results in this study show that the neural network is not significantly more accurate than the linear regression for the data sets employed. The model using differenced data has significantly better results than other models. Among neural network models, the model using differenced data is significantly better than one which simply employs standardized data. The model using factored data also performs somewhat better than that using standardized data, although the difference is not statistically significant. Moreover, the neural network model employing the integrated procedure outperforms the neural network working in isolation.