Fama-French Factor Model: Exposures, Alpha, and Regression Limits

Interpret three- and five-factor regressions; distinguish factor returns, loadings, residuals, and alpha; and validate frequency, universe, statistical uncertainty, and implementation costs.

For educational purposes only; not investment advice. Investing may result in loss.

Direct answer

The Fama-French factor model is a family of empirical asset-pricing models that describes a stock or portfolio’s excess returns using returns on diversified factor portfolios. The three-factor model uses the market, size, and value factors; the five-factor model adds profitability and investment.

The model can help separate raw performance into estimated factor exposure and an unexplained intercept. It does not prove why a premium exists, predict the next winning security, or turn a positive historical alpha into guaranteed manager skill.

How it works

For period t, the three-factor time-series regression is commonly written as:

R i,t - R f,t = alpha i + beta MKT * (R M,t - R f,t) + beta SMB * SMB t + beta HML * HML t + epsilon i,t

The five-factor version adds:

+ beta RMW * RMW t + beta CMA * CMA t

The left side is the asset’s total return minus the same-period risk-free return. MKT-RF is the market excess return. The other factors are zero-investment long-short portfolio returns constructed under published sorting rules:

SMB is Small Minus Big, a size return factor.
HML is High Minus Low book-to-market, a value return factor.
RMW is Robust Minus Weak operating profitability.
CMA is Conservative Minus Aggressive investment.

These labels are not company accounting ratios inserted directly into the regression. Factor definitions vary by model, region, breakpoint, weighting, accounting lag, data vendor, and vintage. For example, the five-factor Data Library’s SMB combines size spreads from book-to-market, profitability, and investment sorts, so it need not equal every other published size factor.

A loading measures return co-movement conditional on the other included factors. A loading of 0.3000 on SMB means a 1.0000 percentage-point SMB return is associated with a 0.3000 percentage-point modeled contribution, holding the other factors fixed. It does not mean the fund holds exactly 30.0000% small-cap stocks.

alpha is the estimated regression intercept at the data frequency. epsilon is each period’s residual. One positive residual is not alpha; alpha is estimated across the sample. Statistical interpretation requires standard errors that address heteroskedasticity and, where relevant, autocorrelation. Economic interpretation also requires fees, trading costs, taxes, capacity, changing exposures, and model choice.

The model can be used descriptively for performance attribution or tested as an asset-pricing model. A high time-series R-squared means the included factors explain much variation in the sample; it does not prove that expected returns are correctly priced or that alpha is skill. Adding factors can reduce an intercept while increasing estimation noise or multicollinearity.

Example

Assume a fund has estimated loadings of 1.0500 on MKT-RF, 0.3000 on SMB, 0.4000 on HML, 0.2000 on RMW, and -0.1000 on CMA. In one month, factor returns are 2.0000%, 1.0000%, -0.5000%, 0.8000%, and 0.3000%, respectively.

The modeled factor contribution is:

(1.0500 * 2.0000%) + (0.3000 * 1.0000%) + (0.4000 * -0.5000%) + (0.2000 * 0.8000%) + (-0.1000 * 0.3000%) = 2.3300%

The components are 2.1000%, 0.3000%, -0.2000%, 0.1600%, and -0.0300%. Suppose the fund’s actual excess return is 2.5000%. Ignoring the estimated intercept for this single-period illustration:

one-month residual = 2.5000% - 2.3300% = 0.1700%

That 0.1700% is one residual, not evidence of persistent alpha. If the risk-free return is 0.3000%, the fund’s total return is 2.8000%; factor regressions must not confuse total and excess returns.

Now assume a regression over 60 months estimates monthly alpha at 0.1500% with a standard error of 0.2000%:

t-statistic = 0.1500% / 0.2000% = 0.7500

An approximate two-sided 95.0000% interval is:

0.1500% +/- (1.9600 * 0.2000%) = -0.2420% to 0.5420% per month

The point estimate is positive but imprecise. Compounding the point estimate alone gives:

annualized alpha point estimate = (1 + 0.1500%)^12 - 1 = 1.8136%

Annualizing does not remove uncertainty and should not be applied mechanically if exposures or alpha vary through time.

Finally, suppose the same fund is compared with a three-factor model and estimated alpha is 0.4000% per month. The five-factor estimate of 0.1500% does not prove that the extra 0.2500 percentage points was misreported performance or that the five-factor model is true. It shows that profitability and investment exposures explain more of the return under the expanded specification.

Risks and verification checklist

Choose the model first: State three-factor, five-factor, regional, global, momentum-augmented, or another specification.
Identify the factor source: Record provider, URL, release date, data vintage, and construction methodology.
Match geography: Use factors appropriate to the portfolio’s investable market and currency.
Match frequency: Align daily, weekly, or monthly asset, factor, and risk-free observations exactly.
Use total returns: Include distributions consistently for the asset, benchmark, and factor portfolios.
Subtract risk-free once: Regress excess return on excess-market and factor returns without double subtraction.
Check units: Confirm decimal versus percent inputs before interpreting coefficients or alpha.
Align timestamps: Prevent accounting information, holdings, or revised data from entering before availability.
Review factor definitions: Inspect breakpoints, weighting, accounting lags, eligibility, and rebalancing.
Use adequate history: Short samples produce unstable loadings, alpha, and standard errors.
Test coefficient stability: Use rolling windows, holdings analysis, and regime splits to detect style drift.
Inspect correlations: Multicollinearity can make individual factor coefficients unstable.
Use robust inference: Address heteroskedasticity, autocorrelation, overlapping returns, and non-normal residuals.
Report uncertainty: Provide standard errors, t-statistics, confidence intervals, and sample size.
Separate residual from alpha: Do not label one favorable period or cumulative residual as the intercept.
Compare specifications: Test CAPM, three-factor, five-factor, industries, momentum, and relevant alternatives.
Include implementation costs: Deduct fees, spread, impact, turnover, borrow, financing, and taxes as appropriate.
Avoid look-ahead bias: Use constituents, accounting data, and factor files available at each historical date.
Test out of sample: Evaluate another period or market without retuning the model after observing results.
Avoid causal claims: A loading describes conditional co-movement and does not by itself prove an economic mechanism.

Common misconceptions

“SMB, HML, RMW, and CMA are company ratios.” They are returns on constructed long-short portfolios based on sorting rules.
“A positive factor loading guarantees a positive return.” Factor returns can be negative, premiums vary, and implementation has costs.
“A positive monthly residual is alpha.” Alpha is a sample intercept estimated with uncertainty; each observation has a residual.
“High R-squared proves the model is correct.” It measures in-sample variation explained, not causal truth or correct expected-return pricing.
“Factor-adjusted alpha is manager skill.” Omitted factors, changing exposures, data mining, luck, fees, and capacity can produce or erase estimated alpha.

Sources

Fama and French, Common Risk Factors in the Returns on Stocks and Bonds.
Fama and French, A Five-Factor Asset Pricing Model.
Fama and French, International Tests of a Five-Factor Asset Pricing Model.
Kenneth R. French Data Library, Description of Fama/French 3 Factors.
Kenneth R. French Data Library, Description of Fama/French 5 Factors (2x3).
Fama and French, The Capital Asset Pricing Model: Theory and Evidence.