The Fama–French Three-Factor Model

The last lesson left CAPM in pieces. One factor — beta on the market — was supposed to explain why some stocks earn more than others, and it simply didn’t: small companies and cheap (high book-to-market) companies kept beating their betas, decade after decade, all over the world. CAPM wasn’t approximately right with a bit of noise; for the cross-section of stocks it was systematically wrong in two specific, repeatable directions.

In 1992–1993, Eugene Fama and Kenneth French stopped patching CAPM and rebuilt it. Their proposal: keep the market factor, but add two more priced risks — a size factor and a value factor — each captured by a clever long–short portfolio you can actually build and trade. The result, the Fama–French three-factor model (FF3), became the default benchmark for “did this fund actually do anything?” for the next thirty years. This lesson builds it from the ground up: the equation, how the two new factors are physically constructed from a sort of every stock on the exchange, how big the rewards are (and how brutally unreliable), and how to read the factor fingerprint of any fund.

Adding two factors to the market

Analogy. CAPM tried to explain every stock’s reward with a single dial: how much it moves with the market. Imagine trying to predict a car’s fuel cost using only its speed. Better than nothing — but you’re ignoring the car’s weight and how often it climbs hills. Fama and French added two more dials. The market dial stays, and beside it sit a size dial (is this a small company or a giant?) and a value dial (is it cheap relative to its accounting net worth, or expensive?). Three dials explain the cross-section far better than one.

Definition. The FF3 model says a stock’s excess return (its return minus the risk-free rate) is explained by three factors plus a residual:

$R_i - R_f = \alpha_i + \beta_i\,(R_m - R_f) + s_i\,\text{SMB} + h_i\,\text{HML} + \varepsilon_i.$

Reading the pieces:

• $R_i - R_f$ is the stock’s return above the risk-free rate (the excess return).
• $R_m - R_f$ is the market factor — the excess return of the whole market, exactly as in CAPM. Its loading is $\beta_i$.
• SMB is Small Minus Big — the size factor, the return of small-cap stocks minus large-cap stocks. Its loading is $s_i$.
• HML is High Minus Low — the value factor, the return of high book-to-market (cheap) stocks minus low book-to-market (expensive) stocks. Its loading is $h_i$.
• $\alpha_i$ is the intercept — the average return left unexplained by all three factors. This is the new, much stricter alpha.
• $\varepsilon_i$ is the random, zero-mean noise.

Why this raises the bar on alpha. Under CAPM, anything you couldn’t pin on market beta got crowned as alpha. FF3 hands two more suspects a chance to take the credit before alpha gets it. A manager who looked brilliant under CAPM — “you beat your beta!” — often turns out to have simply tilted toward small, cheap stocks. Once SMB and HML absorb that tilt, the leftover alpha shrinks, frequently to nothing. Adding factors makes alpha a higher bar to clear, because there are now more ways to explain a return away as ordinary risk exposure rather than skill.

Building SMB and HML: the 2×3 sort

A factor isn’t an abstract idea — it’s a portfolio with a return you can compute every month. Fama and French build SMB and HML from a single annual sort of every stock, and the recipe is worth knowing exactly, because the mechanics explain the factors’ quirks.

The recipe, every June:

1. Sort by size into 2 groups. Rank all stocks by market capitalization and split at the NYSE median into Small (S) and Big (B). (Using the NYSE median, not the all-stock median, prevents the thousands of tiny stocks from dominating the cutoff.)
2. Sort by book-to-market into 3 groups. Independently rank all stocks by book-to-market and split at the NYSE 30th and 70th percentiles into Low (L), Neutral (N), and High (H).
3. Intersect. Crossing 2 size groups with 3 value groups gives 6 value-weighted portfolios: SL, SN, SH, BL, BN, BH.

Analogy — the muffin tin. Picture a muffin tin with 2 rows (Small on top, Big on the bottom) and 3 columns (Low, Neutral, High book-to-market, left to right). Every stock drops into exactly one of the 6 cups based on its size and its cheapness. Now the two factors are just differences between parts of the tin:

SMB (size) = the top row minus the bottom row, averaging across the value columns:

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

HML (value) = the right column minus the left column, averaging across the size rows:

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

Long–short and self-financing. Each factor buys one set of portfolios and short-sells another of equal value. SMB is long all the small portfolios and short all the big ones; HML is long the cheap portfolios and short the expensive ones. Because the long and short legs are roughly equal in dollars, the factor is self-financing (the short proceeds fund the long side) and approximately dollar-neutral and market-neutral — it isolates the size or value spread, not the market’s overall direction. That’s the whole point: a factor return should measure the reward for the tilt itself, stripped of how the market happened to do.

Why the sorts cancel each other out

Here’s the subtle part that makes the design elegant. SMB and HML are both built from the same 6 portfolios, yet they’re meant to measure two different things — size and value — without contaminating each other. The averaging is what pulls that off.

Look again at the formulas. SMB averages across the three value columns (it uses SL, SN, and SH on the small side; BL, BN, and BH on the big side). By including low, neutral, and high book-to-market stocks equally on both the long and short legs, SMB’s value exposure roughly cancels out — whatever value tilt is in the small leg is matched in the big leg. What’s left is (mostly) pure size.

Symmetrically, HML averages across the two size rows (it uses both SH and BH on the long side, both SL and BL on the short side). Small and big appear equally on both legs, so HML’s size exposure roughly cancels, leaving (mostly) pure value.

This is the intuition behind orthogonalization: by constructing each factor so the other characteristic is balanced across its long and short legs, Fama and French make SMB and HML approximately independent. A regression can then attribute a stock’s return to size and value separately instead of tangling them together. It’s not perfect — the factors still have a modest negative correlation in the data — but the averaging is a deliberate, effective purification step, not an accident.

How big are the premia — and how shaky?

A factor is only worth modeling if it actually pays. So how much have the three premia historically earned, and how reliable is that reward? The honest answer is: the rewards are real on a very long horizon, but they are noisy, contested, and wildly regime-dependent — and pretending otherwise is the single biggest mistake in factor investing.

Rough long-run US figures (annualized, and deliberately given as ranges, because they wobble with the sample):

• Market (MKT-RF): about 6–8% per year. The biggest and most dependable of the three.
• Size (SMB): about 2–3% per year — but weak and contested, especially after 1980. The size effect shrank sharply after the original papers were published; much of the early “small stocks win” result was concentrated in tiny, illiquid micro-caps and in January.
• Value (HML): about 3–4% per year over the full sample — but it endured a severe value drought from roughly 2007 to 2020, where HML was flat to negative for over a decade as growth and tech stocks crushed cheap ones. It staged a partial rebound from 2021 onward.

The reliability problem, stated bluntly. Over the full ~1927–2024 sample, the t-statistics on these premia are only about 2 to 3 — barely clearing the conventional “statistically significant” line, and that’s with nearly a century of data. The annual standard deviation of each factor is roughly 10–15%, which utterly dwarfs the 2–4% mean. When the noise is several times the signal, a decade-long negative stretch is not a sign the factor is broken — it’s exactly what you should expect from a positive-but-noisy series. The value drought felt like a death, but it sits comfortably inside the normal range of a factor with that mean and that volatility.

Reading a fund’s factor fingerprint

Once you can estimate a fund’s three loadings — $\beta$ on the market, $s$ on SMB, $h$ on HML (you’ll do this by regression in a later lesson) — you can read its factor fingerprint: a compact description of what kinds of bets the fund really makes.

• $s > 0$ (positive SMB loading) → a small-cap tilt: the fund leans toward smaller companies.
• $s < 0$ → a large-cap tilt.
• $h > 0$ (positive HML loading) → a value tilt: it favors cheap, high book-to-market stocks.
• $h < 0$ → a growth tilt: it favors expensive, glamorous stocks.

A classic small-cap value fund shows up as $s > 0$ and $h > 0$. A large-cap growth fund (think a tech-heavy index) shows up as $s < 0$ and $h < 0$. The sizes of the loadings tell you how hard the fund tilts.

Worked example — decomposing a fund’s return. A small-cap value fund has these estimated loadings, and we’ll use the long-run premia from the chart above:

Add the three contributions to get the return FF3 expects from this fund’s exposures alone:

$\text{Expected excess} = 1.0(6\%) + 0.6(2.5\%) + 0.4(3.5\%) = 6\% + 1.5\% + 1.4\% = 8.9\%.$

Now suppose the fund actually delivered 9.4% in excess return. The FF3 alpha is whatever the factors can’t explain:

$\alpha = 9.4\% - 8.9\% = +0.5\%.$

So of the fund’s 9.4% excess return, 8.9 points were just factor tilts — beta you could have rented through cheap index products — and only 0.5 points were genuine, unexplained skill. A fund that looked like a 9.4-point market-beater is, honestly measured, a 0.5-point one. That gap is the entire reason FF3 exists.

Putting it together

The Fama–French three-factor model keeps CAPM’s market factor and adds two more priced risks — size (SMB) and value (HML) — so a stock’s excess return reads $R_i - R_f = \alpha_i + \beta_i(R_m - R_f) + s_i\,\text{SMB} + h_i\,\text{HML} + \varepsilon_i$. The two new factors are built from a 2×3 sort every June (2 size groups by the NYSE median, 3 book-to-market groups by the 30th/70th NYSE percentiles), giving 6 value-weighted portfolios; SMB is the small portfolios minus the big ones and HML the high book-to-market minus the low. Averaging each factor across the other characteristic makes them approximately orthogonal. The premia are real over long horizons (market ~6–8%, SMB ~2–3% and contested, HML ~3–4%) but noisy and regime-dependent, with a punishing 2007–2020 value drought that was statistically normal, not a death. Estimate a fund’s three loadings and you can read its fingerprint and decompose its return — almost always finding that the headline outperformance was mostly rentable factor beta, with only a thin sliver of true alpha left over.

Next up — Momentum and the five-factor model — we add the strongest and scariest factor of all (momentum, UMD, with its rare but catastrophic crashes) plus the two newer Fama–French additions, profitability (RMW) and investment (CMA), and meet the q-factor model that challenges them. FF3 was the foundation; the modern factor zoo is built on top of it.