Noisy Factors

Abstract

The Fama-French factors are ubiquitous in empirical finance, industry, and law. We find that factor returns differ substantially depending on when the data were downloaded. The effects of these retroactive changes are large. Holding the sample period constant and varying only factor vintages, we show this in three contexts. First, in cross-sectional equity pricing, unconditional alphas of a third of long-short ‘anomaly’ portfolios lose statistical significance. Second, in mutual fund analyses, we show that annual alphas of almost half of individual funds and even portfolios of funds change by more than 1%. Third, in the context of testing asset pricing models, F-statistics from GRS tests of the three-factor model on standard test portfolios vary by up to 40% due only to changes in factor vintages. Our results do not suggest that any particular factor vintage is dominant but point to a source of latent noise that is ignored in conventional tests. We provide suggestions to empiricists on dealing with the noise. Our findings have significant implications for the replicability and robustness of finance research and have a direct bearing on a variety of empirical contexts.

Keywords: Fama French factors, asset pricing, performance evaluation, equities, mutual funds, model fit

JEL Classification: G10, G12, G14, G20, G31