The Fama-French Factor Models

The Fama-French models are the empirical grammar of equity research. When a manager reports "alpha," the working assumption on any modern desk is that it has been measured net of Fama-French factors, beating the market is not enough if a cheap value-tilt portfolio would have done the same. Knowing exactly how the factors are built and tested is table stakes for a quant.

From CAPM to multiple factors

The CAPM says a single factor, the market, prices everything:

$\mathbb{E}[r_i] - r_f = \beta_i\big(\mathbb{E}[r_m] - r_f\big).$

Empirically it fails: small stocks and value (high book-to-market) stocks earn far more than their market betas justify, persistent, economically large CAPM alphas. Fama and French (1993) resolved this not by fixing the theory but by adding factors that span those returns. The three-factor model regresses an asset's excess return on the market plus two long-short portfolios:

$r_{i,t} - r_{f,t} = \alpha_i + \beta_i\big(r_{m,t}-r_{f,t}\big) + s_i\,\text{SMB}_t + h_i\,\text{HML}_t + \varepsilon_{i,t}.$

• SMB (Small Minus Big): the return of small-cap minus large-cap portfolios, the size premium.
• HML (High Minus Low): the return of high book-to-market (value) minus low (growth) portfolios, the value premium. See The Value Factor.

The slopes $s_i, h_i$ are the asset's exposures; the intercept $\alpha_i$ is what the model cannot explain. A well-specified model drives $\alpha_i$ to zero for all test assets.

The 2×3 construction

The factors are built from independent double sorts, and the exact recipe matters. Each June, stocks are split into two size groups (Small/Big) by the NYSE median market cap, and independently into three value groups by the 30th/70th percentiles of book-to-market (Low/Neutral/High). The six intersections give six value-weighted portfolios (SL, SN, SH, BL, BN, BH), and

$\text{SMB} = \tfrac{1}{3}(\text{SL}+\text{SN}+\text{SH}) - \tfrac{1}{3}(\text{BL}+\text{BN}+\text{BH}),$

$\text{HML} = \tfrac{1}{2}(\text{SH}+\text{BH}) - \tfrac{1}{2}(\text{SL}+\text{BL}).$

Notice HML is averaged across size groups so it is (roughly) size-neutral, and SMB is averaged across value groups so it is value-neutral, the double sort orthogonalizes the factors against each other by construction. This is the template every later factor copies.

The five-factor model

Fama and French (2015) added two factors motivated by the dividend-discount identity, a firm's valuation implies a link between expected return, profitability, and investment:

• RMW (Robust Minus Weak): high-profitability minus low-profitability firms (operating profitability). See The Quality Factor.
• CMA (Conservative Minus Aggressive): low-investment (conservative) minus high-investment (aggressive) firms.

$r_{i,t}-r_{f,t} = \alpha_i + \beta_i \text{MKT}_t + s_i \text{SMB}_t + h_i \text{HML}_t + r_i \text{RMW}_t + c_i \text{CMA}_t + \varepsilon_{i,t}.$

A striking finding: once RMW and CMA are included, HML becomes redundant in US data, its average return is absorbed by the profitability and investment factors, because cheap stocks tend to be low-investment and the value premium overlaps with those exposures. Carhart's momentum factor (WML/UMD, winners minus losers) is often added as a fourth/sixth factor because none of the Fama-French factors explain Momentum.

Testing the model: the GRS statistic

The formal test of an asset-pricing model is whether the intercepts from time-series regressions on $N$ test portfolios are jointly zero: $H_0: \alpha_1 = \cdots = \alpha_N = 0$. The Gibbons-Ross-Shanken (GRS) statistic tests exactly this. With $T$ months, $N$ test assets, $K$ factors, and $\hat\Sigma$ the residual covariance,

$\text{GRS} = \frac{T - N - K}{N}\left(1 + \bar f^\top \hat\Omega^{-1}\bar f\right)^{-1} \hat\alpha^\top \hat\Sigma^{-1}\hat\alpha \;\sim\; F_{N,\,T-N-K},$

where $\bar f$ and $\hat\Omega$ are the factor mean vector and covariance. Intuitively GRS asks whether the test-asset alphas are large relative to their estimation error, geometrically, whether the factors' tangency portfolio can be improved by adding the test assets. A model "wins" when GRS fails to reject (small alphas). The competition between three- and five-factor specifications is largely a contest over which produces the smallest GRS across standard test sets.

Worked example: reading a regression

You regress a mutual fund's monthly excess returns on the three factors and get $\hat\beta = 1.0$, $\hat s = 0.4$, $\hat h = 0.6$, $\hat\alpha = 0.10\%$/month with $t_\alpha = 0.8$. Interpretation: the fund is fully market-exposed, tilts toward small caps ($s>0$) and value ($h>0$), and its 1.2%/year "outperformance" is entirely explained by those tilts, the alpha is statistically indistinguishable from zero. The manager was not adding skill; they were selling you a value-and-size portfolio at active fees. This single regression is why factor models transformed performance evaluation: they price the style so you only pay for what is left over.

Failure modes

• Look-ahead in book equity. Book value is known with a lag; Fama-French deliberately lag accounting data 6 months. Skipping this is a classic backtest bug.
• The size factor is fragile. SMB's premium is weak, concentrated in tiny illiquid stocks and in January, and largely vanishes after controlling for quality, many now treat size as a conditioning variable, not a standalone premium.
• Factor definitions are not unique. Book-to-market vs. other value measures, operating vs. gross profitability, results move with definitions, so replication demands the exact recipe.
• Regional and regime instability. HML's redundancy is a US, recent-sample result; it does not hold everywhere or always (The Value Factor's lost decade).

In interviews

Expect to write the three-factor regression, name SMB and HML and how they are constructed from the 2×3 sort, and extend to the five-factor model (RMW, CMA) with the one-line intuition that profitability and investment come from the valuation identity. The sharp follow-ups: "What does a significant alpha in this regression mean?", return the model cannot explain, i.e., skill or a missing factor, and "Why add momentum separately?", because Fama-French factors do not span it, so Carhart's UMD is bolted on. If asked how you'd test the model, cite the GRS test of jointly-zero alphas. See Factor Investing for the broader risk-premium-versus-anomaly framing.