A Performance Comparison of Large-n Factor Estimators

This paper uses simulations to evaluate the performance of various methods for estimating factor returns in an approximate factor model when the cross-sectional sample (n) is large relative to the time-series sample (T). We study the performance of the estimators under a variety of alternative speci?cations of the underlying factor structure. We ?nd that 1) all of the estimators perform well, even when they do not accommodate the form of heteroskedasticity present in the data; 2) for the sample sizes considered here, accommodating heteroskedasticity does not deteriorate performance much when simple forms of heteroskedaticity are present; 3) estimators that handle missing data by substituting ?tted returns from the factor model converge to the true factors more slowly than the other estimators.