Guide

Fama-French three-factor model explained

Harbor Capital's small-cap value sleeve beat the Russell 2000 by 280 basis points annualized over five years. The PM called it stock-picking edge. A risk analyst ran a Fama-French three-factor (FF3) regression and found something else: loadings of 0.92 on SMB (small minus big) and 0.71 on HML (high minus low book-to-market) explained almost all of the outperformance. Residual Jensen's alpha was statistically indistinguishable from zero. The sleeve was not broken — but its label was wrong. Investors were paying active fees for factor exposure they could replicate with two ETFs.

Eugene Fama and Kenneth French published the three-factor model in 1992 as a direct response to a flaw in the Capital Asset Pricing Model (CAPM): market beta alone could not explain why small stocks and cheap (high book-to-market) stocks earned higher average returns. FF3 adds two systematic risk factors — size and value — so that most style tilts show up as factor loadings rather than mysterious alpha. This guide defines the three factors, walks through the regression equation, explains how academic factor portfolios are built, interprets betas and residual alpha, covers the Harbor attribution refactor, compares FF3 to CAPM and the later five-factor extension, provides a model decision table, common pitfalls, and a production checklist alongside our broader factor investing guide.

Why CAPM left returns on the table

CAPM says expected excess return is proportional to market beta:

E[R_i − R_f] = β_i (E[R_m − R_f])

Empirically, portfolios sorted on beta did not line up cleanly with average returns. Worse, two persistent anomalies survived every CAPM test: small-cap stocks outperformed large caps on average (the size effect), and value stocks (high book-to-market ratio) outperformed growth stocks (low book-to-market). A manager tilted toward small value could look brilliant in a CAPM regression — high alpha — when in fact they were harvesting well-known risk premia.

Fama and French proposed that size and value are additional dimensions of systematic risk. Investors who hold small or distressed firms earn compensation for bearing liquidity constraints, financial distress, and informational uncertainty that the average investor avoids. Whether you call that “risk” or “mispricing” is still debated; for portfolio attribution, FF3 is the standard vocabulary either way.

The three factors: market, SMB, and HML

FF3 decomposes stock or portfolio returns into three factor exposures plus a residual:

R_i − R_f = α_i + β_i(R_m − R_f) + s_i SMB + h_i HML + ε_i

Market (MKT-RF): Excess return of the broad equity market over the risk-free rate. This is CAPM's market factor; every diversified equity portfolio loads heavily on it.
SMB (Small Minus Big): Return of a diversified small-cap portfolio minus a diversified large-cap portfolio, both value-weighted within their size groups. Positive s_i means the asset behaves like small caps; negative means large-cap tilt.
HML (High Minus Low): Return of high book-to-market (value) stocks minus low book-to-market (growth) stocks. Positive h_i is a value tilt; negative is growth.

Interpreting the regression output

Run a time-series regression of the portfolio's monthly excess returns on the three factor returns (Ken French publishes free factor data at Dartmouth). The coefficients tell you how much of each factor the portfolio embeds:

β_i ≈ 1 with s ≈ 0, h ≈ 0 — market-like, cap-neutral, style-neutral.
s > 0, h > 0 — small-cap value profile (Harbor's sleeve).
α_i — average return not explained by the three factors after controlling for loadings. This is Jensen's alpha in a multi-factor world.

Standard errors on alpha matter. A 150 bp alpha with a 200 bp standard error is noise; allocators should demand statistical significance and economic significance after fees.

How factor portfolios are constructed

Academic SMB and HML series are not arbitrary ETFs. French's methodology (updated annually) roughly follows:

Universe: All NYSE, AMEX, and NASDAQ common stocks with required accounting data.
Size breakpoint: NYSE median market cap splits small vs big. NASDAQ small stocks are included in the small group (important detail — NASDAQ-heavy portfolios load differently on SMB).
Value sort: Book equity divided by market equity (B/M) at fiscal year end; 30th and 70th percentile breakpoints define growth, neutral, and value ter tiles.
Six portfolios: Small Value (SV), Small Neutral (SN), Small Growth (SG), Big Value (BV), Big Neutral (BN), Big Growth (BG) — value-weighted within each cell.
Factor returns: SMB = (SV + SN + SG)/3 − (BV + BN + BG)/3; HML = (SV + BV)/2 − (SG + BG)/2.

Commercial index providers (Russell, MSCI, CRSP) use different rules, so an “FF3 loading” on a live fund will never match the academic series exactly. For due diligence, use the same factor data source consistently across managers and document any proxy mismatch.

From loadings to expected return

If you accept FF3 as a pricing model (strong form), expected excess return is:

E[R_i − R_f] = β_i E[MKT-RF] + s_i E[SMB] + h_i E[HML]

Historical factor premia (US, 1927–present in French data) are approximately 8% market, 2% SMB, and 4% HML annualized — but premia vary wildly by decade. Value crushed growth in the 2000s; growth dominated the 2010s. Size has been weak in the US since the mid-1980s. Using full-sample averages to forecast the next decade is a common mistake.

In practice, FF3 is used more for attribution than forecasting: decompose realized return into “you owned the market,” “you were small,” “you were value,” and “what's left.” That decomposition feeds performance attribution and manager selection alongside information ratio and tracking error.

Harbor Capital refactor: relabeling the sleeve

After the FF3 regression, Harbor's allocator team rebuilt the small-cap value sleeve review around factor truth rather than narrative:

Replicate the tilt passively. A 50/50 blend of a small-cap ETF and a value ETF reproduced 94% of the sleeve's factor exposure at 12 bps vs 85 bps active fee.
Reframe alpha threshold. Residual alpha needed to exceed fee differential (~73 bps) plus a significance buffer. Observed alpha was −8 bps (insignificant).
Split the mandate. Core allocation moved to passive factor ETFs; a smaller true-alpha sleeve was sized only where information coefficient on proprietary signals cleared internal hurdles.
Report FF3 loadings quarterly. Investors now see β, s, h, and rolling 36-month alpha with confidence bands — not just benchmark-relative return.

The PM kept their job; the fund saved fees and set honest expectations. That is what a good FF3 workflow looks like in production.

FF3 vs CAPM vs five-factor vs smart beta

Model	Factors	Best for	Limitation
CAPM	Market beta only	Quick beta estimates, regulatory cost of equity, simple dashboards	Misses size/value tilts; inflates alpha for style funds
FF3	Market + SMB + HML	Equity manager attribution, style fund due diligence, academic baseline	Ignores profitability and investment; weak on momentum
FF5 (2015)	FF3 + RMW + CMA	Growth/value nuance, quality/profitability tilts, modern factor ETFs	More factors = more estimation noise; data requirements grow
Smart beta ETFs	Rules-based single-factor products	Cheap, transparent factor access for allocators	Index rules differ from French factors; factor crowding and cycles
Custom multi-factor	Firm-specific signals + optimization	Quant shops with proprietary breadth and IC	Overfitting risk; backtest ≠ live FF3 loadings

Common pitfalls

Confusing factor exposure with skill — high raw return on a small-value fund is often SMB + HML, not alpha. Always regress before praising managers.
Using the wrong factor data — French US vs developed ex-US vs Europe series differ. A global fund needs global factors or explicit proxy notes.
Short regression windows — 12 months of monthly data gives unstable loadings. Use 36–60 months for allocation decisions; note structural breaks after mandate changes.
Ignoring multicollinearity — SMB and HML correlate (~0.3–0.5 historically). Individual stock loadings are noisy; portfolio-level regressions are more reliable.
Assuming constant premia — factor returns are cyclical. Value had a lost decade; size has been episodic. Attribution is backward-looking.
Book-to-market vs other value metrics — HML uses B/M. P/E, EV/EBITDA, or cash-flow yield tilts are related but not identical; loadings on HML will understate some value strategies.
Survivorship in custom backtests — when building in-house factor portfolios, delisted stocks matter enormously for true SMB/HML replication.

Production checklist

Download consistent Ken French (or equivalent) factor data for your universe.
Regress portfolio excess returns on MKT-RF, SMB, HML monthly; store coefficients.
Report β, s, h, alpha, and R² with standard errors.
Use 36+ months of data unless the mandate changed (then split samples).
Compare implied factor exposure to holdings-based attribution for sanity.
Benchmark active fee against passive factor replication cost.
Document proxy differences if using commercial ETFs instead of academic factors.
Extend to FF5 when profitability/investment tilts are material.
Pair FF3 output with tracking error and information ratio for allocator reports.
Re-run regressions after major style drift or manager turnover.

Key takeaways

FF3 adds size (SMB) and value (HML) factors to CAPM so style tilts appear as loadings, not fake alpha.
The regression equation is the standard tool for equity manager attribution and fund due diligence.
Academic factor portfolios use NYSE breakpoints and B/M sorts; live ETFs approximate but do not match exactly.
Residual alpha after FF3 must beat fees and survive statistical tests to justify active management.
Factor premia are cyclical; FF3 explains the past better than it forecasts the future.