1.3.4 Cross-Sectional Factor Models

Basic Prerequisites: 03 · Factor Construction and Linear Algebra & Matrices (inverse matrices, projections, GLS).

---

1. Intuition & Practical Objective

The Fama–French factors of the previous pages are macro-style traded spreads. Commercial portfolio risk models (MSCI Barra, Axioma) flip the regression cross-sectionally: they treat the asset characteristics (B/M, earnings yield, size, momentum, volatility) as the known exposures $X$, and estimate the factor returns $f_t$ by regressing, at each time step, the whole cross-section of asset returns on those exposures. The practical objective: produce (a) a realized factor return for each style each period, (b) a factor covariance matrix for risk, and (c) a portfolio-level risk decomposition - the machinery a risk system runs daily.

The two regression directions, contrasted:

	Time-series (FF)	Cross-sectional (Barra/Axioma)
Regress	asset returns on factor returns	cross-section of asset returns on exposures
Known	$f_t$ (traded factor portfolios)	$X$ (standardized characteristics)
Estimated	loadings ${\beta }_i$ per asset	factor returns $f_t$ per period
Used for	performance attribution, alpha	risk model, factor P&L, portfolio attribution

---

2. Mathematical Ground Truth & Derivations

2.1 The cross-sectional factor model (Tsay §9.3.1, eqs. 9.1–9.4)

Each period, for all $N$ assets:

$$r_t=X_t{f}_t+u_t,$$

where $r_t\in {\mathbb{R}}^N$ is the vector of returns, $X_t\in {\mathbb{R}}^{N\times K}$ the matrix of standardized factor exposures (z-scores of B/M, momentum, earnings yield, size, etc.), $f_t\in {\mathbb{R}}^K$ the realized factor returns, and $u_t$ the asset-specific residuals with $\text{Cov}(u_t)=D=\text{diag}\{{\sigma }_1^2,\dots ,{\sigma }_N^2\}$ (the specific-risk matrix).

2.2 WLS/GLS estimation - the two-step Barra recipe (Tsay eq. 9.7–9.8)

Because residuals have unequal variance, use weighted least squares with weights $w_i=1/{\sigma }_i^2$ (diagonal inverse specific-risk):

$${\hat{f}}_t=(X_t^{\top }{D}^{-1}{X}_t{)}^{-1}{X}_t^{\top }{D}^{-1}{r}_t.$$

Equivalently this is GLS with $V=D$. In practice Barra does a two-step OLS → GLS refinement (Tsay eq. 9.8): estimate OLS factor returns, use the residuals to estimate specific variances ${\sigma }_i^2$, then re-run with the inverse-specific-risk weights.

Factor-mimicking portfolio (Tsay §9.3.1). The estimator is linear in returns, ${\hat{f}}_t={\omega }^{\top }{r}_t$, with

$$\omega =(X^{\top }{D}^{-1}X{)}^{-1}{X}^{\top }{D}^{-1},$$

an $N\times K$ matrix whose $k$-th column is the portfolio of assets that mimics the $k$-th factor’s return - long high-exposure names, short low-exposure ones, optimally tilted by inverse specific risk.

Industry special case. If a block of $X$ is industry dummies, the corresponding factor return is (up to weighting) the industry-mean return - OLS recovers exactly the industry-average return (Tsay §9.3.1).

2.3 Covariance factorization - the risk model (Tsay eq. 9.3)

Given factor covariance $\Omega =\text{Cov}(f_t)$ and specific-risk diagonal $D$, the asset covariance is

$$\Sigma =X\Omega X^{\top }+D.$$

This is the entire point of the model: instead of estimating the $\frac{N(N-1)}{2}$ pairwise covariances (ill-conditioned when $N$ is large), estimate the $K\times K$ factor covariance $\Omega$ plus $N$ specific variances. A portfolio’s factor risk is then $\text{Var}(w^{\top }r)=w^{\top }X\Omega X^{\top }w+w^{\top }Dw$.

2.4 Fama–MacBeth second pass (cross-sectional risk premiums)

The FF cross-sectional pass (FF1992) regresses each period’s returns on the estimated loadings to recover factor risk premiums: monthly regressions $r_{it}={\lambda }_{0t}+{\sum }_j{\lambda }_{jt}{\beta }_{ij}+{ϵ}_{it}$, then average the slopes over time, ${\hat{\lambda }}_j=\frac{1}{T}{\sum }_t{\hat{\lambda }}_{jt}$, with $t$-statistic ${\hat{\lambda }}_j/s({\hat{\lambda }}_j)$. This is what produced FF1992’s headline: size slope $-0.15\%$/month ($t=-2.58$) and beta slope only 0.46 standard errors from zero.

---

3. Computational Implementation - one cross-section of Barra-style WLS + covariance count

Stdlib only. Simulate $N=150$ assets with known exposures and true factor returns, estimate the factor returns by WLS (inverse-specific-risk weights), and demonstrate the covariance-dimension reduction.

The WLS recovers all three factor returns near their true values (the residual error is the specific-risk noise), and the covariance factorization collapses 11,175 pairwise covariances to just 156 parameters - a $\sim$72× reduction, which is why a factor risk model is numerically tractable where a full $N\times N$ covariance is not.

---

4. Failure Modes & First-Principles Breakdowns

1. Exposure quality is everything. The cross-sectional model assumes $X$ (standardized fundamentals) is measured correctly and point-in-time. Restatements, stale book values, and look-ahead in the exposures directly corrupt ${\hat{f}}_t$ (see Accounting Quality & Red Flags).
2. Multicollinear exposures. If several exposure columns are near-linear (five value variants), $X^{\top }{D}^{-1}X$ is near-singular and individual factor returns explode while the fit is fine - the same mechanism as 05 · Failure Modes.
3. Specific-risk misspecification. If $D$ is wrong (residuals are not independent or the variances are stale), WLS is suboptimal and the factor covariance $\Omega$ is biased - the risk model understates portfolio tail risk exactly when markets stress.
4. Factor-return estimation is unstable under regime shifts. Cross-sectional betas and factor P&L rotate with the market; a model fit in one regime misprices the next (see Kalman Filtering for time-varying exposure estimation).

---

5. References

• Tsay, Analysis of Financial Time Series
• Fama & French, “The Cross-Section of Expected Stock Returns” (JF, 1992)
• MSCI Barra / Axioma risk-model documentation
• Hastie et al., The Elements of Statistical Learning

---

6. Connected Graph Bridges

• Back: 03 · Factor Construction · Index Hub
• Forward: 05 · Failure Modes & Practice
• Downstream risk: Portfolio Optimization (factor-risk-budgeted portfolios) · Quantitative Risk (the factor covariance is the risk model)
• Base: Linear Algebra & Matrices · Econometrics & Time Series