The Economic Value of Nonlinear Predictions in Asset Allocation

Abstract

Predictions of asset returns and volatilities are heavily discussed and analyzed in the finance research literature. In this paper, we compare linear and nonlinear predictions for stock- and bond index returns and their covariance matrix. We show in-sample and out-of-sample prediction accuracy as well as their impact on asset allocation results for short-horizon investors. Our data comprises returns from the German DAX stock market index and the REXP bond market index as well as their joint covariance matrix over the period 01/1988 - 12/2007.

The comparison of a linear and nonlinear prediction approach is the focus of this study. The results show that while out-of-sample prediction accuracies are weak in terms of statistical significance, asset allocation performances based on linear predictions result in significant Jensen's alpha measures and Sharpe-ratio and are further improved by nonlinear predictions.

Keywords: nonlinear prediction, neural networks, asset allocation

JEL Classification: C32, C45, C53, G11