Approximate Quantile Factor Models

Weakly dependent idiosyncratic errors are standard in mean factor models, but quantile factor models have been analyzed only under independence. The mean-case definition of weak dependence, a bound on the eigenvalues of the error covariance matrix, does not carry over: I show that it neither implies nor is implied by the property that controls quantile estimation, since uncorrelated errors can share a pervasive quantile component. I instead define approximate quantile factor models through a concentration condition on the idiosyncratic array. Under it, the quantile factor analysis (QFA) estimator is consistent at a rate that deteriorates by the square root of a dependence constant, which is minimax optimal, and the number of factors is consistently estimated. Dependence changes the asymptotic variance, for which I provide HAC and thresholded estimators. Monte Carlo experiments confirm these predictions. In FRED-QD and FRED-MD, standard errors ignoring dependence understate uncertainty, most strongly in the tails.