Title:Regularizing Bayesian Predictive Regressions

Abstract:We develop a novel approach for regularizing Bayesian predictive regressions with applications in empirical finance. We regularize both expected returns and variance- covariance matrices in vector auto-regression (VAR) and seemingly unrelated regression (SUR) models. To illustrate our methodology, we analyze an asset allocation problem due to Barberis (2000) and Kandel and Stanmbaugh (1996). We exploit a duality between regularization penalties and prior distributions to reinterpret their Bayesian analysis as a regularization procedure. This leads to computational advantages and a full regularization path to the illustrate sensitivity to prior assumptions. Rather than marginalizing out prior hyper-parameters, we use a predictive cross-validation approach to select the optimal amount of regularization. We show how our regularized estimates can be exploited in a sequential portfolio allocation context. Finally, we conclude with directions for future research.