Incremental Alpha of a New Signal

A researcher brings you a new signal with an information coefficient of 0.05 — a respectable number, worth publishing on its own. Should it go in the book? The answer depends entirely on a question the raw IC can't answer: how correlated is this signal with the ten signals already running? A brand-new 0.05-IC signal that's 90% correlated with an existing factor adds almost nothing to the portfolio's actual forecasting power, no matter how good it looks in isolation. What matters is not the signal's alpha — it's the alpha it adds on top of what's already there.

Isolating what's actually new

The standard way to measure this is to regress the new signal on the existing ones and keep only the residual — the part of the new signal that existing signals can't already predict:

$s_{\text{new}}^{\perp} = s_{\text{new}} - \big(\hat\gamma_0 + \hat\gamma_1 s_1 + \hat\gamma_2 s_2 + \dots\big)$

In words: fit the new signal as a linear combination of the signals already in the book, then subtract that fitted value out. What's left, $s_{\text{new}}^{\perp}$, is orthogonal to everything already running — genuinely new information, by construction. The incremental IC is then this residual signal's correlation with forward returns, not the raw signal's.

Worked example. A desk already trades a value signal and a quality signal. A researcher proposes a new "cheap and improving margins" signal with a raw IC of 0.05. Regressing the new signal against value and quality gives an $R^2$ of 0.64 — 64% of the new signal's cross-sectional variation is already explained by the two existing signals, meaning its correlation with the existing combination is $\sqrt{0.64} = 0.80$. Computing the residual signal and correlating that with forward returns gives an incremental IC of 0.018 — nearly two-thirds of the raw 0.05 evaporates once you account for overlap. The signal isn't worthless (0.018 is still positive, still real), but it's a much smaller addition than the headline number implied, and whether it clears the bar for the operational cost of adding a new data feed and a new line in the risk model is a much closer call at 0.018 than at 0.05.

This scatter shows what a high-overlap signal pair looks like: at $\rho = 0.8$, knowing one signal already tells you most of what you need to know about the other. Drag the correlation down toward zero and picture the new signal's residual — that's the version of the signal actually worth adding, and it's a much thinner, noisier cloud than the raw correlation with the existing factor suggested.

Why this matters more than raw IC

Grinold and Kahn's fundamental law of active management ties a strategy's information ratio to its IC and the breadth of independent bets it makes — $IR \approx IC \times \sqrt{\text{breadth}}$. Adding a signal that's 80% correlated with an existing one barely increases breadth, because the two signals are mostly voting on the same set of names for the same reasons. Adding a signal that's genuinely uncorrelated with what's already there, even with a smaller raw IC, increases breadth more and often adds more to the book's overall information ratio than a bigger, more redundant IC would.

What this means in practice

Research teams that skip this step end up with bloated signal libraries: forty "distinct" signals that are really six independent ideas wearing different names, each one adding data costs, model complexity, and turnover without adding much forecasting power. It also creates a subtler risk: several near-duplicate signals independently telling the portfolio optimizer to load up on the same names produces concentrated, correlated positions that look diversified on paper (forty signals!) but behave like one large bet in a drawdown — a milder version of the crowding that hits stat arb books when everyone's "independent" signals turn out to agree.

In interviews

Be ready to write the orthogonalization regression from memory and explain in one sentence why raw IC is the wrong number to report when a signal is being added to an existing book. If asked how you'd screen a stream of proposed signals, describe regressing each new candidate against the current signal set and ranking candidates by incremental IC, not raw IC — and note that a small, genuinely orthogonal signal can be more valuable to the book than a larger, redundant one, because of how breadth enters the fundamental law of active management.