Is the recent Fama-French augmentation of their classic three-factor (market, size, book-to-market) model of stock returns with profitability and investment factors a major advance? In their November 2016 paper entitled “Five Concerns with the Five-Factor Model”, David Blitz, Matthias Hanauer, Milan Vidojevic and Pim van Vliet identify five concerns regarding the five-factor model. Based on empirical and theoretical (rationale) grounds, they note that:
- The five-factor model retains the capital asset pricing model (CAPM) assumption that higher market beta for a stock means higher return. Mounting evidence indicates that the relationship is not positive but flat, or even negative.
- The five-factor model ignores the strong and pervasive momentum factor.
- The new profitability and investment factors: (a) fail to explain relationships between stock returns and closely related variables (some of which appear to be better new factors); and, (b) appear to be absent in other asset classes and in stocks before 1970.
- While the three-factor model relies on reward-for-risk explanations, justifications for including profitability and investment factors are unclear (opening the door for empirical factor set snooping).
- It seems unlikely that the five-factor model will to settle main asset pricing debates or lead to consensus.
In summary, evidence indicates that the Fama-French five-factor model is not a major advance in explaining stock return behaviors.
Other cautions regarding factor models of stock returns include:
- Factor models generally assume linear relationships between stock returns and factors. Relationships may not be linear.
- Factor models generally estimate premiums based on gross returns. Accounting for factor portfolio periodic reformation frictions, which can vary materially by stock, may produce different (and more realistic) premiums.
- Factor set snooping, both direct and inherited from other studies, without accounting for the associated statistical bias (probability of discovery factor sets that are lucky in available data) produces models that overstate predictive power.
