Thesis (Ph. D.)--University of Rochester. William E. Simon Graduate School of Business Administration, 2017.
Lewellen (2015), building on the prior work of Haugen and Baker (1996) and
Hanna and Ready (2005), showed that Fama-MacBeth regression slopes with
many anomaly variables can be used to forecast returns out-of-sample. This paper
proposes a different way to "combine" the anomaly variables, using the minimum
distance estimator that is more efficient than the Fama-MacBeth. The method
essentially weights period-by-period slopes by their estimated precisions. By substantially reducing the amount of noise in the estimates, this method allows a
trading strategy that produces larger long-short portfolio spreads and alphas than
those produced by the Fama-MacBeth method, when stocks are sorted by the resulting fitted values. In direct comparisons, it is also shown that such a strategy
generates significant abnormal returns not spanned by the returns from the Fama-
MacBeth strategy, and that the explanatory power of these return estimates is
larger than that of the Fama-MacBeth estimates. The results are robust to different variable selections, time periods and rolling window lengths. The strategy
also performs better while having lower transaction costs. In addition, I present
an application that uses my method to generate the level, slope and curve factor
model in the spirit of Clarke (2016), and show that such a three-factor model performs substantially better than his version of the factor model and also favorably to other leading factor models.