Many mortgage reform proposals suggest replacing Fannie Mae and Freddie Mac (the GSEs) with private entities. A common assumption underlying these proposals is that unlike the GSEs, private insurers will properly manage risk and set fair prices. Inconsistent with this assumption, this paper presents evidence that private insurers less effectively managed home price risks during the 2000-2006 housing boom than the GSEs did. Mortgage origination data reveal that the GSEs were selecting loans with increasingly higher percentages of down payments, or lower loan to value ratios (LTVs), in boom areas than in other areas. These lower LTVs in boom areas reduced the GSEs' exposure to overheated markets. Furthermore, the decline of LTVs in boom areas stems entirely from the segment insured by the GSEs only, and none of the decline stems from the segment where private mortgage insurers take the first loss position. Private insurers also did not lower their exposure to home price risks along other dimensions, including the percentage of high LTV GSE loans they insured and the percentage of insured mortgage balance. My results highlight that post-crisis reform of the mortgage insurance industry should carefully consider additional factors besides moral hazard induced by the government guarantees, such as mortgage insurers' future home price assumptions and the industry organization of the mortgage origination chain.
Estimation of Quadratic Forms for High Dimensional Time Series: Correcting Finite Sample Bias in the Mean Variance Frontier with Alexander Aue and Debashis Paul Presented at: NBER-NSF Time Series Conference 2016, Columbia University; UC Berkeley Risk Management Seminar; University of Waterloo, Department of Statistics and Actuarial Science In finite samples, risk is known to be underestimated in the mean-variance frontier. This paper proposes a novel algorithm correcting this finite sample bias when asset returns follow a high dimensional time series. Assuming a linear process formulation studied in Liu, Aue, and Paul (2015), an algorithm is proposed to estimate the spectral distributions of the coefficient matrices of the linear process by making use of the asymptotic behavior of the empirical spectral distributions of symmetrized autocovariance matrices. This leads to the formulation of a strategy for the estimation of the mean-variance frontier, utilizing the estimates of the coefficient matrix spectra. The proposed method is extended to a setting in which the returns are assumed to have a factor model structure with observed factors, while the unknown idiosyncratic terms are assumed to belong to the aforementioned class of linear processes. The performance of the proposed methods is examined through extensive simulation studies.
JOURNAL PUBLICATIONS On Marcenko-Pastur Law for Linear Time Series, Annals of Statistics, vol. 43, no. 2, pp. 675-712, 2015. with Debashis Paul and Alexander Aue Presented at: NBER-NSF Time Series Conference 2013, Federal Reserve Board; USC Marshall School of Business, Data Sciences and Operations