Nonparametric prediction of stock returns with generated bond yields

The question of whether empirical models are able to forecast the equity premium more accurately than the simple historical mean is intensively debated in the nancial literature. The low prediction power is disappointing, even when using nonparametric models that make use of typical predictor variables. Motivated by the co-movement of bond and stock returns, the inclusion of the current bond yield in a prediction model is proposed. This results in a notable improvement in the prediction of stock returns, as measured by the validated R2. Since the current bond yield is unknown, it is predicted in a prior step. The essential point is that the inclusion of the generated bond can be seen as a kind of dimension and complexity reduction that imposes more structure in an appropriate way that circumvents the curse of dimensionality and complexity.