Doctor of Philosophy (PhD)



Document Type



Portfolios of mortgage loans played an important role in the Great Recession and continue to compose a material part of bank assets. The distribution of mortgage portfolio returns, and consequently, the risk of these portfolios, is quite distinct even from other fixed income asset classes. This dissertation contains three essays, each aiming to analyze a specific component of risk in mortgage portfolios and role of geographical diversification in reducing this risk.

The first essay investigates how cross-sectional dependence in the underlying properties flows through to the loan returns, and thus, the risk of the portfolio. In addition to demonstrating this relationship theoretically, this essay demonstrates how the spatially dependent structure of the underlying housing returns is revealed in the mortgage market by a shock to the default rate. The resulting increase in the asset correlations reduces the effectiveness of any geographical diversification present in the portfolio.

Even when the distribution of mortgage returns is known, the ability to reduce portfolio risk through geographical diversification can be limited due to the concentration of mortgage debt in major metropolitan areas. The second essay aims to model this geographical concentration for various partitions of the mortgage market and examine the role this has on limiting investors' ability to diversify risk. This is accomplished by fitting the empirical regularity from regional science known as the rank-size rule to measure this concentration.

The third and final essay in this dissertation focuses on modeling the mortgage default decision and imputing unobserved factors that may bias the estimated impact of observed factors such as the loan-to-value ratio. As alluded to in the first essay, the default rate, or probability of default ex-ante, is an important determinant of the observed correlation across mortgage returns. This essay develops a ridge regression model, which is tuned to maximize out-of-sample predictive performance using cross-validation, that imputes these unobserved factors while preventing model overfitting.

Committee Chair

Pace, R. Kelley