Forecasting financial time series using multiple Kernel Learning

Abstract

This thesis introduces a forecasting procedure based on Multiple Kernel Learning to predict and measure the influence of several economic variables in the process of predicting the equity premium of the S&P 500 Index. In the experiments of Welch and Goyal they determined that, using linear models, those economic variables had an unreliable effect on the predictive capabilities of the models. The experiments performed in this thesis with MKL use the same data in an attempt to predict with non-linear models. The kernels that are part of the MKL procedure are multivariate dynamic kernels adapted for time series. The presented financial variables have a questionable impact on the predictive capabilities of the developed models due to the data being noisy. Some of the kernel methods for time series may not be able to extract any relevant information from exogenous variables, as they are matched in the results by a simple RBF kernel. MKL shows a poor capacity at selecting the best combination of kernels as it is also matched by RBF, and even the kernels that MKL uses. However the experimental results show that the presented methods have better predictive capabilities than the linear models.