Pure Factor Portfolios

Sort stocks by book-to-market and go long cheap, short expensive, and you get a value characteristic portfolio, but cheap stocks also tend to be smaller and more volatile than expensive ones. Some of that portfolio's return is really a small-cap effect or a volatility effect leaking in through the back door, not pure value at all. A pure factor portfolio is engineered to strip that contamination out entirely: unit exposure to the one factor you want, exactly zero exposure to every other factor in the model.

The analogy

Imagine measuring how much a car's top speed depends on engine size, but every big-engine car in your sample also happens to be lighter and more aerodynamic. A naive comparison of big-engine versus small-engine cars mixes all three effects together. A careful engineer instead builds two matched cars, identical weight, identical aerodynamics, differing only in engine size, and compares those. That matched pair isolates engine size alone. A pure factor portfolio does the equivalent matching across an entire universe of stocks at once, using regression instead of hand-picking pairs.

Building the portfolio

Take the cross-sectional regression from before, stock returns on $K$ factor exposures $X$ (an $N \times K$ matrix, one row per stock, one column per characteristic):

$R = X f + \epsilon, \qquad \hat{f} = (X^\top X)^{-1} X^\top R.$

The weight matrix $W = (X^\top X)^{-1} X^\top$ is exactly the set of pure factor portfolios: row $k$ of $W$ gives the weights on every stock for the portfolio that isolates factor $k$. In words: this is the same least-squares machine that produces $\hat{f}$, the estimated factor return, but read the weights it implicitly used rather than just the final coefficient, and those weights define an actual, holdable portfolio. By construction, that portfolio has exposure exactly 1 to factor $k$ and exposure exactly 0 to every other factor in $X$, because that is precisely what makes $(X^\top X)^{-1}X^\top X$ come out to the identity matrix.

Think of $(X^\top X)^{-1}$ as "un-mixing" correlated characteristics the same way this transform can be inverted to recover an original, unrotated shape, the pure factor weights undo the correlation between characteristics so each resulting portfolio responds to only one axis.

Worked example 1: why a simple sort is contaminated

Five stocks have book-to-market $2.0, 1.5, 1.0, 0.7, 0.3$ and size scores $-1.5, -0.8, 0.0, 0.9, 1.6$ (negative means small). Notice the pattern: cheap stocks (high book-to-market) also tend to be small here, correlation is strongly negative. A simple long-cheap-short-expensive characteristic portfolio, ignoring size entirely, is long stocks that are both cheap and small, and short stocks both expensive and large. Its return mixes the value premium with the size premium in unknown proportions.

Worked example 2: purging the contamination

Running the two-factor cross-sectional regression $R_i = f_{\text{value}}\, \beta_{\text{val},i} + f_{\text{size}}\, \beta_{\text{size},i} + \epsilon_i$ on those five stocks (rather than a single-characteristic sort) produces a weight vector for the value row of $W = (X^\top X)^{-1}X^\top$ that, unlike the simple sort, gives negative weight to some of the small-cheap stocks to cancel out their size contribution, and positive weight to large-cheap stocks to compensate. The resulting pure value portfolio, by construction, has a measured size exposure of exactly 0.00 when you multiply its weights by the size scores, versus the simple sort's uncontrolled leakage. The tradeoff: the pure factor portfolio typically holds more extreme long and short positions, and more of them, than the simple sort, since it must actively work against the correlation between characteristics.

What this means in practice

Commercial risk models (Barra, Axioma) report pure factor returns, not characteristic-portfolio returns, in their standard factor exposure reports, precisely so that a "value return of 2% this month" means value alone, not value contaminated by whatever else happened to correlate with it that month.