Residualise the Signal or Constrain the Portfolio?

Once a researcher decides a signal should not bet on, say, sector or size, there are two genuinely different ways to enforce that, and it is easy to reach for whichever one is more convenient without noticing they are not interchangeable.

Two places to intervene

Residualising the signal happens at the research stage, before portfolio construction even begins. The raw signal is regressed against the unwanted exposure (sector membership, size, beta, whatever), and only the leftover — the residual, the part of the signal not explained by that exposure — is carried forward as the score used to rank and weight stocks. By the time the portfolio-construction step sees the signal, the unwanted exposure is already gone from the data it's working with.

Constraining the portfolio leaves the raw signal untouched and instead adds a rule to the portfolio optimiser: the final book must have zero (or near-zero) net exposure to sector, size, beta, or whatever the constraint targets. The optimiser is free to use the raw, un-cleaned signal to rank stocks, but it has to find weights that respect the constraint when it assembles the actual positions.

Why the results differ

Residualising changes what "high signal" means before any portfolio math happens — a stock that was mostly attractive because it's a small-cap now scores lower, full stop, regardless of what else is in the book. Constraining leaves that stock looking just as attractive as before; the optimiser instead compensates by pairing it with an offsetting position elsewhere, which can mean holding a position for no reason other than to cancel out someone else's exposure — a trade that exists purely for hedging, not conviction.

This matters most when the optimiser has limited room to manoeuvre. With a small universe or tight risk budget, a hard portfolio constraint can force the optimiser into an unnatural, low-conviction book just to satisfy it, whereas residualising spreads the same adjustment gently across every stock's score and lets normal portfolio construction proceed from there.

	Residualise the signal	Constrain the portfolio
When it acts	Before ranking, in the research pipeline	At portfolio construction, after signal is fixed
What changes	The score itself	The weights, given the unchanged score
Risk	Can overcorrect and remove real information correlated with the exposure	Can force offsetting positions with no independent conviction
Easiest to audit	Yes — one transformed number per stock	Harder — the effect is spread across the whole optimisation