The Grinold Alpha Formula

A model outputs a score: this stock ranks 1.4 standard deviations above the universe average on the signal. A position-sizing engine needs an expected return, in percent, to decide how large a bet that deserves. The Grinold alpha formula is the standard bridge between the two — a score, on its own arbitrary scale, becomes a real, unit-consistent forecast of return.

The formula

$\alpha_i = IC \cdot \sigma_i \cdot z_i$

In words: the expected return $\alpha_i$ for stock $i$ equals three things multiplied together. $IC$ is the information coefficient — the correlation, measured historically, between the signal's scores and the returns that followed. $\sigma_i$ is stock $i$'s own return volatility — a signal implies a bigger dollar-and-cents return swing for a volatile stock than a calm one, for the same rank. $z_i$ is the stock's standardised score on the signal (how many standard deviations above or below the universe average it ranks).

Each piece answers a distinct question. $IC$ answers "how good is this signal, on average, across the whole universe and history?" $z_i$ answers "how strongly does this signal favour this specific stock, right now?" $\sigma_i$ answers "how much should that conviction translate into an actual percentage return, given how much this stock moves?" Multiply the three and you get a name-specific, properly-scaled forecast, built from numbers a researcher can actually estimate.

A worked example

A signal has a measured monthly IC of 0.04, estimated from several years of decile-spread testing. Stock A ranks $z = 2.0$ (strongly favoured) and has annualised return volatility of 35%, translating to monthly volatility $\sigma \approx 35\% / \sqrt{12} \approx 10.1\%$. Stock B ranks $z = 2.0$ as well, but is a calmer name with annualised volatility of 18%, or $\sigma \approx 5.2\%$ monthly.

Even though both stocks get the identical score from the model, their Grinold-formula forecasts differ:

• Stock A: $\alpha = 0.04 \times 10.1\% \times 2.0 \approx 0.81\%$
• Stock B: $\alpha = 0.04 \times 5.2\% \times 2.0 \approx 0.42\%$

Nearly double the expected return for the same signal score, purely because Stock A is more volatile — a stock that moves more, moves more in both directions, so the same relative conviction implies a bigger absolute return. This is precisely the adjustment a position-sizing step needs and a raw model score cannot provide on its own.

	Stock A	Stock B
Signal score $z$	2.0	2.0
Monthly volatility $\sigma$	10.1%	5.2%
$IC$	0.04	0.04
Grinold expected return $\alpha$	0.81%	0.42%

Where it's used, and where it breaks

The formula is the standard last step of a forecasting pipeline, immediately before position sizing — it's the point where "researcher's model output" becomes "the number a portfolio construction step consumes." It also underlies capacity and risk-budgeting conversations, since expressing every signal as an expected return in common units is what lets a PM compare an equity signal to a rates signal on equal footing.

It breaks down in two specific ways. First, $IC$ is a population-average number; using a single global $IC$ for every name assumes the signal works equally well across the whole universe, which is rarely exactly true — some desks fit a separate $IC$ per sector or liquidity bucket. Second, the formula assumes scores and returns are linearly related, which is a reasonable approximation near the middle of the distribution and a worse one at the extremes, where very high scores often don't correspond to proportionally higher realised returns.