Date of Award

Summer 8-15-2017

Author's School

Graduate School of Arts and Sciences

Author's Department

Business Administration

Additional Affiliations

Olin Business School

Degree Name

Doctor of Philosophy (PhD)

Degree Type



In the first chapter, we present a Bayesian framework to search for a parsimonious subset of given factors to explain the cross-section of expected returns. The framework relies on two important assumptions, that factors are traded portfolio excess returns or return spreads and the stochastic discount factor is linear in the factors. The framework uses the Bayesian marginal likelihood criterion to jointly compare all possible linear factor pricing models to find the one that is best supported by the evidence. This marginal likelihood comparison requires proper priors which we devise in a creative way, taking account of the large dimension of the parameter space and the large dimension of the model space. The framework overall is self-contained and can be used with minimum user intervention. We provide detailed simulation evidence about the high accuracy of the method to locate the true model among a large collection of models. This accuracy is shown to increase with sample size as per the asymptotic theory of the marginal likelihood. Finally, we provide detailed results from the application of our framework to the actual data on 13 common risk factors. In this real-data analysis, we use a training sample that runs from January 1968 to December 1972 to form our prior distribution and an estimation sample that runs from January 1973 to December 2015 to estimate and compare 40,955 factor models. This analysis reveals that the highest marginal likelihood model is a student-t distributed factor model with 4 degrees of freedom and 8 pricing factors.

In the second chapter, we point out that the marginal likelihood based method in Barillas and Shanken (2017a) is flawed because of its reliance on improper priors and, consequently, the empirical findings in the paper about the set of pricing factors supported by the data are without merit. Extensive simulation exercises shows that our method developed in Chapter one performs sharply better than the method of Barillas and Shanken (2017a).

In the third chapter, we develop an easily implemented Bayesian strategy for improved forecasts of the market excess returns. This strategy relies on 1) the use of multiple predictors in the predictive regression; 2) zero lower bound constraints on the Bayesian predictive mean (the mean marginalized over the parameters); and 3) conjugate prior distributions for smooth sequential updating and calculation of needed posterior quantities. This strategy produces striking results in the prediction of the market excess returns and outperforms the results from 8 other predictive methods, such as univariate and model combined univariate predictive regressions. We show that with a set of 11 common market return predictors used in the literature, and for an investor with power utility, the utility gain from our strategy exceeds that from the other predictive methods by a wide margin.


English (en)

Chair and Committee

Siddhartha Chib

Committee Members

Guofu Zhou, Matthew Ringgenberg, Asaf Manela, Werner Ploberger,


Permanent URL:

Available for download on Wednesday, July 19, 2119