Feature Neutralization

Suppose a feature meant to capture a firm's earnings quality turns out, on inspection, to just be a proxy for which sector a stock is in — tech firms always score high, utilities always score low. A model trained on that raw feature isn't learning anything about earnings quality; it's rediscovering "is this a tech stock," which you already knew and could have modeled directly and more robustly. Feature neutralization is the fix: regress the feature (or the model's prediction) against known exposures you don't want it riding on, and keep only what's left over — the residual.

The mechanics: regress out, keep the residual

Neutralization is ordinary linear regression used as a filter rather than a predictor. Fit the feature $x$ against a set of exposures $Z$ you want to remove (sector dummies, market beta, size, or any prior factor):

$$x = Z\beta + \varepsilon$$

Every symbol here: $x$ is the raw feature for each stock, $Z$ is a matrix of the unwanted exposures, $\beta$ is the fitted sensitivity of $x$ to those exposures, and $\varepsilon$ is the residual — what's left of $x$ after its relationship with $Z$ is subtracted out. You then use $\varepsilon$, not $x$, as the model's input. In plain English: you're asking "of this feature's variation, how much is just riding on things I already understand, and how much is genuinely new information?" — and keeping only the new information.

Worked example: neutralizing a value score against sector

A "value score" feature has these raw values across five stocks, three of which are tech and two of which are utilities:

Stock	Sector	Raw score	Sector average	Neutralized
A	Tech	0.8	0.7	0.1
B	Tech	0.6	0.7	−0.1
C	Tech	0.7	0.7	0.0
D	Utility	0.3	0.35	−0.05
E	Utility	0.4	0.35	0.05

Here neutralization is done the simple way — subtract each stock's sector mean — which is exactly what a regression on sector dummies alone would produce. Notice the raw score makes it look like tech stocks (0.6–0.8) beat utilities (0.3–0.4) outright, but that gap is entirely explained by sector. The neutralized column shows the real, sector-independent signal: stock A stands out within tech, and stock E stands out within utilities, which is the comparison a cross-sector model actually needs to learn from.

What this means in practice

Neutralization is standard practice in quant equity research precisely because so many promising-looking features turn out to be disguised sector, size, or momentum bets — regressing those out before feeding a feature to a model (or before evaluating a factor's returns) forces an honest answer to "does this add anything new." It's also used on the output side: neutralizing a model's final predictions against known factors before portfolio construction, so the resulting book doesn't inherit unintended sector or beta tilts.