Arbitrage Pricing Theory and Multi-Factor Models › Estimating Factor Exposures

With the factor model framework established, we now turn to the practical task of estimating an asset's factor exposures. As we discussed in Part III's chapter on Regression Analysis, time-series regression provides a natural estimation approach.

For an individual stock or portfolio, we regress excess returns on the factor returns:

$$R_{i,t} - r_{f,t} = \alpha_i + \beta_i^{MKT} \cdot MKT_t + \beta_i^{SMB} \cdot SMB_t + \beta_i^{HML} \cdot HML_t + \epsilon_{i,t}$$

where:

• $R_{i,t} - r_{f,t}$: excess return on asset $i$ at time $t$
• $\alpha_i$: estimated alpha
• $\beta_i^{FACTOR}$: estimated factor loadings
• $FACTOR_t$: factor return at time $t$
• $\epsilon_{i,t}$: residual at time $t$

This regression has a natural interpretation. The dependent variable, excess return, is what we seek to explain. The independent variables, the factor returns, represent the systematic risk sources. The regression coefficients tell us how much the asset's return moves, on average, in response to each factor.

The regression coefficients $\beta_i^{MKT}$, $\beta_i^{SMB}$, and $\beta_i^{HML}$ are the estimated factor loadings. The intercept $\alpha_i$ represents the average return not explained by the factors: positive alpha suggests outperformance, negative alpha suggests underperformance.

The interpretation of factor loadings is straightforward:

• $\beta^{SMB} > 0$: The asset behaves like a small-cap stock
• $\beta^{SMB} < 0$: The asset behaves like a large-cap stock
• $\beta^{HML} > 0$: The asset behaves like a value stock
• $\beta^{HML} < 0$: The asset behaves like a growth stock

These interpretations connect statistical estimates to economic meaning. A stock with a large positive SMB loading tends to rise when small stocks outperform large stocks, regardless of the company's actual market capitalization. The factor loading captures behavioral similarity rather than category membership.

An alternative approach estimates factor risk premia rather than individual factor loadings. In cross-sectional regression, we use known or estimated betas as explanatory variables and current returns as the dependent variable:

$$R_{i,t} = \lambda_{0,t} + \lambda_{MKT,t} \hat{\beta}_i^{MKT} + \lambda_{SMB,t} \hat{\beta}_i^{SMB} + \lambda_{HML,t} \hat{\beta}_i^{HML} + \nu_{i,t}$$

where:

• $R_{i,t}$: return on asset $i$ at time $t$
• $\lambda_{0,t}$: intercept (zero-beta rate) at time $t$
• $\lambda_{FACTOR,t}$: estimated risk premium for the factor at time $t$
• $\hat{\beta}_i^{FACTOR}$: estimated factor loading for asset $i$ (from first pass)
• $\nu_{i,t}$: pricing error

This regression is run across all assets at each point in time, yielding time series of estimated risk premia $\lambda_t$. The average of these estimates gives the historical factor risk premium.

The logic of cross-sectional regression differs fundamentally from time-series regression. In time-series regression, we ask: "Given that we know what the factors did, how sensitive was this particular asset?" In cross-sectional regression, we ask: "Given that we know each asset's sensitivities, what premium did the market pay for each type of risk?"

The Fama-MacBeth procedure (1973) formalizes this two-step approach:

1. First pass: Time-series regressions to estimate each asset's factor betas
2. Second pass: Cross-sectional regressions at each date to estimate factor premia

This methodology remains the standard for testing factor pricing models. It provides point estimates of factor premia and standard errors that account for the time-series variation in estimated premia.