Structural Risk Models

1 The return decomposition

The model is built in four steps:

1. Section (§2.1). Read today's exposures \(\mathbf X_t\) off company data — no returns involved.
2. Section (§2.2). For each past period \(t\), regress that period's realized returns for each stock (\(\mathbf r_t\)) on the exposures (\(\mathbf X_{t-1}\)) known before it. The coefficients are the factor returns \(\hat{\mathbf b}\), one set per period \(t\). The section then shows how the history of \(\hat{\mathbf b}\) gives \(\mathbf F\), and the leftover residuals give \(\boldsymbol\Delta\). The same section shows that this estimator is itself a set of portfolios, one per factor.
3. Section (§2.3). Steps 1 and 2 produce the three objects needed to assemble the covariance matrix: \(\hat{\mathbf V}_{t+1} = \mathbf X_t\mathbf F\mathbf X_t^{T} + \boldsymbol\Delta\).
4. Section (§3). Split a portfolio's risk using this.

The period subscripts are dropped below wherever only one period is in play.

Part 2 — Building the model

2.1 Building the Exposure Matrix \(\mathbf X\)

Every entry of \(\mathbf X_{t-1}\) has to be measurable before period \(t\) begins, which rules out anything derived from the returns it is meant to explain. The columns come in two kinds.

Industry columns. \(X_{nk} = 1\) if asset \(n\) is in industry \(k\), else 0; conglomerates get fractional membership by sales or assets, and the row still sums to one. These absorb the co-movement that comes from doing the same business.

Risk-index columns. Continuous descriptors — size, value, momentum, volatility, leverage, liquidity, growth. Raw descriptors are in incompatible units (log market cap against a book-to-price ratio), so each is standardized across the estimation universe:

The Barra convention takes \(\mu_k\) cap-weighted and \(\sigma_k\) equal-weighted. The two choices do different jobs, and the cap-weighted \(\mu_k\) is doing something exact. Let \(w_n\) be the benchmark's cap weights, so \(\sum_n w_n = 1\), and let \(\mu_k\) be defined with those same weights:

Then the benchmark's own exposure to risk index \(k\) is

So it is true by construction, and only for the one portfolio whose weights were used to centre the descriptor. An equal-weighted \(\mu_k\) would leave a non-zero benchmark exposure, and so would cap weights from a different index.

That zero is what makes the exposures readable: a portfolio's exposure is directly a bet against the benchmark, and \(\mathbf x = \mathbf X^{T}\mathbf h\) needs no re-centering before it is interpreted. The equal-weighted \(\sigma_k\) keeps a handful of megacaps from setting the scale, so a one-unit exposure means one cross-sectional standard deviation across the universe rather than across the index's largest members.

Four practical points, none of which appear in the algebra but all of which decide whether the fitted model is any good:

• Winsorization. The cross-sectional regression is least squares, so descriptors are standardized, trimmed at roughly \(\pm 3\) standard deviations, then standardized again.
• Point-in-time data. Exposures must use what was knowable at \(t-1\): as-first-reported fundamentals, not restatements, and filing dates rather than fiscal period ends. Otherwise \(\mathbf X\) is not genuinely known before the period and the backtest inherits look-ahead.
• Composite descriptors. A descriptor is one raw measured quantity; a risk-index column is usually several of them standardized separately and then averaged. Value combines book-to-price with earnings-to-price and cash-flow-to-price; momentum combines returns over several horizons. Each is a noisy proxy for the same underlying quality — book-to-price is distorted by intangibles, earnings-to-price by one-off items — so averaging cancels the idiosyncratic measurement error and keeps what they share.
• Missing data. Assign the peer-group mean rather than dropping the asset, which after standardization is an exposure at or near zero. Dropping assets shrinks the cross-section and biases it toward larger, better-covered names.

2.2 Estimating the Factor Returns \(\mathbf b\)

Fix a period \(t\) and write eq. 1 out one asset at a time, with \(X_{nk}\) the \((n,k)\) entry of \(\mathbf X_{t-1}\):

That is \(N\) equations sharing \(K\) unknowns \(b_{1t},\dots,b_{Kt}\). Every quantity in eq. 4 belongs to the one period \(t\), and the index that runs is \(n\): the observations are assets, not dates. Since \(K \ll N\) the system is overdetermined, so it is solved by least squares, and the whole thing is run again from scratch next period.

Getting that direction backwards is the most common misreading of the model, because the better-known construction runs the other way:

OLS to get \(\mathbf b\). Minimizing \(\mathbf u^{T}\mathbf u = (\mathbf r - \mathbf X\mathbf b)^{T}(\mathbf r - \mathbf X\mathbf b)\) gives

which weights every stock equally. But specific risk is strongly heteroskedastic — a micro-cap's specific volatility runs several times a mega-cap's — so equal weighting lets the noisiest assets in the cross-section drive the estimate. Weighting each observation by its precision \(1/\Delta_{nn}\) instead means minimizing \((\mathbf r - \mathbf X\mathbf b)^{T}\boldsymbol\Delta^{-1}(\mathbf r - \mathbf X\mathbf b)\):

By Gauss–Markov this is the minimum-variance linear unbiased estimator when the residuals are uncorrelated with unequal variances, which is precisely a diagonal \(\boldsymbol\Delta\) with unequal entries — though only with the true \(\boldsymbol\Delta\), since substituting an estimate for it matches the bound only asymptotically. OLS is the special case \(\boldsymbol\Delta = \sigma^{2}\mathbf I\): equal weighting is optimal only if every stock carries the same specific risk, which no equity universe does. In practice \(1/\Delta_{nn}\) is often approximated by market cap or its square root, since specific variance falls roughly with size.

Full column rank. Both estimators invert a \(K\times K\) matrix built from \(\mathbf X\), so \(\mathbf X\) must have rank \(K\) or neither has a unique solution. The industry columns of §2.1 are the standard way this fails: every asset belongs somewhere, so those columns sum to \(\mathbf e\), the vector of ones, and adding a market or intercept column makes \(\mathbf X\) rank-deficient. Two standard fixes: drop one industry, which makes every industry return relative to the omitted one, or keep all of them and impose a constraint such as \(\sum_k c_k b_k = 0\), where \(c_k\) is industry \(k\)'s share of total market capitalization, which keeps the industries symmetric and interprets the intercept as the market return.

The estimator is a set of portfolios

Equation 6 is linear in \(\mathbf r\). Write it as \(\hat{\mathbf b} = \mathbf H^{T}\mathbf r\) and read off the \(N\times K\) matrix that does it:

Column \(k\) of \(\mathbf H\) is a vector of holdings, and \(\hat b_k = \mathbf h_k^{T}\mathbf r\) is that portfolio's return. So each estimated factor return is the realized return on an actual portfolio, not a statistical abstraction. Whatever the factor did this month, some book of long and short positions earned exactly that.

What those portfolios hold is pinned down by one identity:

Entry \((k,l)\) of \(\mathbf H^{T}\mathbf X\) is portfolio \(k\)'s exposure to factor \(l\). So portfolio \(k\) has unit exposure to factor \(k\) and zero exposure to every other factor: it is a pure bet on one factor, hedged against the rest.

What these portfolios minimize. For a single attribute \(\mathbf a\), the minimum-variance portfolio holding one unit of exposure to it — \(\min_{\mathbf h}\mathbf h^{T}\mathbf V\mathbf h\) subject to \(\mathbf h^{T}\mathbf a = 1\) — is called the characteristic portfolio of \(\mathbf a\), and works out to \(\mathbf V^{-1}\mathbf a/(\mathbf a^{T}\mathbf V^{-1}\mathbf a)\). Column \(k\) of eq. 7 solves the same problem with \(\boldsymbol\Delta\) in place of \(\mathbf V\), and \(K\) constraints in place of one:

Appendix A does the Lagrangian; at \(K = 1\) it collapses to \(\boldsymbol\Delta^{-1}\mathbf a/(\mathbf a^{T}\boldsymbol\Delta^{-1}\mathbf a)\), the characteristic-portfolio formula with \(\mathbf V \to \boldsymbol\Delta\).

They are estimation devices, not trades. Nothing in eq. 9 mentions turnover, borrow, position limits, or transaction costs. Factor portfolios typically hold thousands of names at tiny weights, carry large gross exposure relative to net, rebalance completely every period by construction, and take positions in assets no one can short. Tradeable factor products are separate objects, built with those constraints in the optimization.

From the estimates to \(\mathbf F\) and \(\boldsymbol\Delta\)

The circularity. Equation 6 needs \(\boldsymbol\Delta\), but \(\boldsymbol\Delta\) is estimated from the variance of the regression residuals \(\hat{\mathbf u} = \mathbf r - \mathbf X\hat{\mathbf b}\), which needs \(\hat{\mathbf b}\). Two ways out, usually combined: iterate within a period — start from OLS or a size proxy, form residuals, re-weight, repeat until the estimates settle — or use \(\hat{\boldsymbol\Delta}\) from prior periods, which is what production systems do, since a lagged estimate cannot introduce look-ahead and specific variances move slowly.

From the panel to \(\mathbf F\) and \(\boldsymbol\Delta\). Running eq. 6 each month for \(t = 1,\dots,T\) gives a panel \(\{\hat{\mathbf b}_t\}\) of \(K\) time series. Then \(\mathbf F\) is the sample covariance matrix of those series and \(\Delta_{nn}\) is the time-series variance of asset \(n\)'s residuals \(\hat u_{nt}\). Both are ordinary time-series estimates — but now on \(K\) series rather than \(N\), which is the whole reason the exercise is tractable.

Each period's regression used a different \(\mathbf X_{t-1}\), so the panel is only comparable across periods because the descriptors are re-standardized every period by eq. 2. That is what fixes the units: \(\hat b_{tk}\) means the same thing in every \(t\) — the return to a one-cross-sectional-standard-deviation bet on characteristic \(k\) — so covariances taken down the panel are meaningful. Standardization is not cosmetic; without it factor returns from different decades would be denominated differently and \(\mathbf F\) would be an average of incompatible numbers.

\(\hat{\mathbf b}_t\) is an estimate, not the factor return itself, so its sample covariance carries estimation error on top of the true \(\mathbf F\) and is biased upward. Production models correct for this, and typically also shrink \(\mathbf F\) and scale it to match realized volatility.

2.3 The covariance result

Take \(\mathbf b\) and \(\mathbf u\) centered. \(\mathbf V\) is a covariance, which is mean-insensitive. Then \(\mathbf V = \operatorname{Var}(\mathbf r) = \mathbb E[\mathbf r\mathbf r^{T}]\).

Equation 10 read as a forecast. Standing at date \(t\) and wanting next month's covariance matrix, eq. 10 says

\(\mathbf X_t\) determines the exposures that will govern next month's returns based on today's market capitalizations, book-to-price ratios and industry memberships. \(\mathbf F\) and \(\boldsymbol\Delta\) are estimated from history, so they describe factor behaviour in general. \(\hat{\mathbf F}_t\) is \(K\times K\), and the same \(\hat{\mathbf F}_t\) is used for every pair of stocks, so it cannot be what makes one pair differ from another. \(\mathbf X_t\) is \(N\times K\), one row per stock: row \(n\) holds stock \(n\)'s own characteristics, so Apple's row and Pfizer's row are different numbers, and that difference is the only thing separating one covariance from another.

A sample covariance matrix has one input and one only, past returns, with no slot for the fact that a company is now ten times the size it was. Five years a micro-cap and then a tripling leaves sixty monthly observations of which fifty-nine describe the small version of it, and they clear the window one month at a time. Recomputing \(\mathbf X_t\) today moves that stock's size exposure at once. The exception is \(\hat{\boldsymbol\Delta}_t\), one variance per asset taken from that asset's own past residuals, so specific risk lags in exactly the way a sample estimate does.

The subscripts on \(\hat{\mathbf F}_t\) and \(\hat{\boldsymbol\Delta}_t\) mark them as estimated from history through \(t\), and that history was built from the past exposures \(\mathbf X_0,\dots,\mathbf X_{t-1}\) (§2.2). Today's \(\mathbf X_t\) appears in eq. 11 and nowhere else — it is information no historical estimate contains.

Dimension check. \(\mathbf X\) is \(N\times K\), \(\mathbf F\) is \(K\times K\), \(\mathbf X^{T}\) is \(K\times N\), so \(\mathbf X\mathbf F\mathbf X^{T}\) is \(N\times N\) and conformable with \(\boldsymbol\Delta\). Both terms are symmetric, so \(\mathbf V\) is.

Entry by entry.

Because \(\boldsymbol\Delta\) is diagonal the second term vanishes for \(i \ne j\), so two different assets covary only through their factor exposures. Two biotech firms are correlated because both load on the same industry and size factors, not because the model was told anything about the pair. The converse is the restriction: two stocks with identical rows of \(\mathbf X\) get identical covariance with every other asset, and differ only in \(\Delta_{nn}\).

Parameter count. \(\mathbf F\) contributes \(K(K+1)/2\) and diagonal \(\boldsymbol\Delta\) contributes \(N\). Sixty factors over a thousand assets is \(1830 + 1000 = 2830\) parameters against 500,500 — 177 times fewer. The \(NK\) numbers in \(\mathbf X\) are not estimated at all.

Part 3 — Using the model

3 Risk Decomposition

Substituting eq. 10 into portfolio variance and collecting \(\mathbf X^{T}\mathbf h\):

Every attribution report is a reading of eq. 13. The first term depends on the portfolio only through its \(K\) factor exposures \(\mathbf x\), so two portfolios with different holdings and the same exposures carry identical common-factor risk. The second is \(\sum_n h_n^{2}\Delta_{nn}\) for diagonal \(\boldsymbol\Delta\), which shrinks like \(1/N\) as holdings spread out: an equal-weighted book drives specific risk toward zero while its common-factor risk stays put. That is the precise sense in which factor risk is the part diversification cannot remove.

Why §2.2 could minimize over \(\boldsymbol\Delta\). Using \(\boldsymbol\Delta\) rather than \(\mathbf V\) changes nothing, because the constraint has already spent the factor risk:

\(F_{kk}\) is the same number for every feasible \(\mathbf h\), so minimizing total variance and minimizing specific variance are the same problem here and have the same minimizer. \(\boldsymbol\Delta\) appears because GLS weights by the covariance of the regression's errors, and it survives into the portfolio statement because requiring unit exposure to one factor and zero to the rest has already fixed the common-factor variance at \(F_{kk}\).

4 Assumptions

What the construction leans on.

• \(\mathbf X\) is known at \(t-1\) and fixed through the period. Used to pull \(\mathbf X\) out of the expectation in eq. 10. Exposures drift within a month as prices move — a stock's size and momentum exposures at month-end are not the ones the model used.
• \(\operatorname{Cov}(\mathbf b,\mathbf u) = \mathbf 0\). Not directly testable, since both sides are estimated from the same regression, and the residuals are orthogonal to \(\mathbf X\) by construction whether or not the population statement holds. A genuinely omitted factor that loads on the included characteristics violates it.
• \(\boldsymbol\Delta\) diagonal. The strongest assumption in the model. Real residual correlation survives any factor set — dual listings, merger pairs, supply-chain partners, companies sharing a single customer. Concentrated portfolios are where the understatement bites, since eq. 13 then treats correlated bets as independent ones.
• Linearity. Equation 1 admits no interaction or nonlinear terms; an effect that appears only for small and illiquid names has nowhere to go but \(\mathbf u\).
• Stationarity of \(\mathbf F\) and \(\boldsymbol\Delta\) over the estimation window, so that a sample covariance of past \(\hat{\mathbf b}_t\) forecasts the next period. Factor volatilities are visibly regime-dependent, which is why production models rescale.
• \(\boldsymbol\Delta \succ 0\). Needed for \(\mathbf V \succ 0\) and for the GLS weighting in eq. 6. \(\mathbf X\mathbf F\mathbf X^{T}\) alone has rank at most \(K\).
• \(\mathbf X\) has full column rank \(K\). Needed for eqs. 5 and 6 to have unique solutions; the industry-plus-intercept collinearity of §2.1 is the standard way this fails.
• It is a risk model, not a return model. Nothing here estimates \(\mathbb E[\mathbf b]\), and the factor structure carries no claim that exposure is compensated. Using \(\mathbf X\mathbb E[\mathbf b]\) as an alpha is a separate assertion needing separate evidence.

A Appendix — factor portfolios as a constrained minimization

Claim: column \(k\) of \(\mathbf H\) in eq. 7 solves eq. 9. The Lagrangian carries a \(K\times 1\) multiplier \(\boldsymbol\lambda\), one per constraint:

Imposing the constraint on eq. A2 determines \(\boldsymbol\lambda\), and substituting back gives \(\mathbf h\):

which is column \(k\) of eq. 7. The objective is strictly convex and the constraints affine, so this is the unique minimizer, not merely a stationary point.

The \(K = 1\) case. With one factor, \(\mathbf X\) is a single column \(\mathbf a\) and \(\mathbf e_k = 1\). Then \(\mathbf X^{T}\boldsymbol\Delta^{-1}\mathbf X = \mathbf a^{T}\boldsymbol\Delta^{-1}\mathbf a\) is a scalar and eq. A3 reads

The characteristic portfolio of \(\mathbf a\) from §2.2, with \(\boldsymbol\Delta\) wherever \(\mathbf V\) stood. The \(K > 1\) version adds the requirement that the portfolio be neutral to the other \(K-1\) factors.