Bayesian portfolio selection under a multifactor asset return model with predictive model selection

This paper addresses the problem of portfolio selection under a multifactor asset return model, using Bayesian analysis to deal with uncertainties in parameter estimation and model specification. These sources of error are ignored in the classical mean-variance method. We apply two approaches: the empirical Bayes method, and Bayesian model averaging. The previous literature on Bayesian portfolio selection has paid little attention to the researcher's choice of factors contributing to the asset return prediction. This paper uses a previously published criterion to quantify the predictive power of several candidate models and justify this choice. Using data from the US and Japanese stock markets, a comparative analysis is conducted between the two Bayesian methods and the classical mean-variance method. A major finding pertinent to investors is that the influence of each asset return factor varies with time, depending heavily on the state of the market. Both Bayesian methods perform better than the classical method, but the difference between them is not great.