We investigate the short-horizon stock and bond return predictability in a predictive regression and a predictive system using a Bayesian framework. In contrast to the predictive regression where the expected returns are modeled as a linear function of predictors, in the predictive system this assumption is relaxed, and predictors do not account for the entire variance in expected returns. We argue that a fair comparison of these two models has not been drawn yet. In our approach both models are estimated using the same Bayesian methodology and we carefully construct corresponding priors for both models. In particular, we focus on the prior beliefs about the coefficient of determination. By allowing for various distributions of this coefficient
we account for different degrees of optimism about predictability. In our comparative
study we take a look at the models from an investor’s perspective. Therefore, we evaluate the out-of-sample performance of both models by calculating certainty equivalent returns implied by an asset allocation strategy. We show that relaxing the assumption of perfect predictors does not seem to pay offout-of-sample. Furthermore, we find that extreme optimism or pessimism about predictability decreases the performance of both models.