The Factor Alignment Problem

A customs officer trained to spot forged watches confiscates anything that resembles the pattern of a fake — including, occasionally, a genuine antique that happens to share the same telltale scratches. Meanwhile, contraband disguised in a way the training never covered walks straight through, untouched. A portfolio optimizer has exactly this blind spot. It can't distinguish a stock's genuine, skill-based alpha from a risk factor it already recognizes if the two happen to look statistically similar — so it "confiscates" real alpha that correlates with a known factor, shrinking the position as if it were just an unrewarded risk bet, while a real risk the model has no factor for passes through completely unchecked.

Two models, two different views of the same stocks

An alpha model ranks stocks by expected return using its own signals; a risk model estimates variance using its own factors, built independently, usually by a different team on a different schedule. The optimizer combines them by maximizing expected return for a given level of risk:

$$\max_{\mathbf{w}} \; \mathbf{w}^\top \boldsymbol{\alpha} - \lambda\, \mathbf{w}^\top \boldsymbol{\Sigma}\, \mathbf{w} .$$

In words: pick weights $\mathbf{w}$ that make the portfolio's expected alpha $\mathbf{w}^\top \boldsymbol{\alpha}$ as large as possible, penalized by $\lambda$ times the portfolio's variance $\mathbf{w}^\top \boldsymbol{\Sigma} \mathbf{w}$ from the risk model's covariance matrix $\boldsymbol{\Sigma}$. The optimizer has no way of knowing why $\boldsymbol{\alpha}$ is high on a given stock — if the alpha signal happens to be correlated with, say, the risk model's Value factor, the optimizer sees a position that raises both expected return and measured risk together, and it will size it more cautiously than the alpha alone would justify, even if the underlying alpha has nothing to do with Value at all.

Worked example 1 — real alpha, mistaken for a risk bet

Suppose a stock-selection signal, built from analyzing supply-chain data, ranks a stock highly. That stock also happens to have a high loading on the risk model's Value factor, purely by coincidence of which companies show up cheap this quarter. The optimizer, penalizing variance, treats part of the position's expected return as compensation for taking a known, already-diversifiable Value bet rather than paying for it as pure alpha — so instead of the position size the raw alpha signal would justify, the optimizer scales it down, say from an unconstrained $8m position to $5m, specifically because the risk model is "double-counting" the Value-correlated portion as risk rather than skill.

Worked example 2 — a real risk, invisible to the model

Now the reverse case: an alpha model likes five semiconductor-equipment suppliers that all depend on a single customer's capital-spending cycle — a genuine shared risk factor, but not one the risk model's standard sector and style factors capture, since the model only sees "Semiconductors" as a broad sector, not this narrower customer-concentration risk. The optimizer, seeing five names it believes are only correlated at the ordinary sector level, happily sizes all five near their full unconstrained weight, judging the combination "diversified." If that shared customer cuts spending, all five names fall together — a risk that was real all along but invisible to the model that was supposed to catch it.

What this means in practice

Quant desks address this by building an "alpha alignment factor" — effectively adding the alpha signal itself into the risk model as a recognized factor, so the optimizer stops penalizing it as if it were unexplained risk — or by regularly reconciling the alpha and risk teams' factor definitions so neither team is working blind to what the other considers a risk. Left unaddressed, a firm can end up systematically under-sizing its best ideas and over-sizing bets on risks nobody was tracking.