Rank IC Versus Pearson IC

The information coefficient (IC) measures how well a signal's cross-sectional values line up with subsequent returns (see information coefficient). "Line up" can be measured two ways: ordinary Pearson correlation between raw signal values and raw returns, or Spearman rank correlation between their ranks. These usually tell a similar story, but when they disagree it's rarely coincidence — it's usually a handful of outlier stocks distorting the Pearson version in a way the rank version ignores by design.

An analogy: finish order versus stopwatch gaps

Imagine judging how well a coach's pre-race predictions matched actual results. One way: compare predicted times to actual times directly — but if one runner had a freak injury and finished dramatically slower, that single gap can dominate the comparison, saying little about whether the coach judged the order of everyone else correctly. The other way: compare only the rankings, ignoring time gaps entirely. Pearson IC is the first method, thrown off by one extreme value; Rank IC is the second, caring only about ordering — usually the actual thing an investment signal is used for, since portfolio construction overweights and underweights stocks relative to each other rather than betting on exact return magnitudes.

The two measures, one symbol at a time

Pearson IC is the ordinary linear correlation between a signal's raw cross-sectional values $s_i$ and subsequent raw returns $r_i$ across stocks $i = 1, \ldots, n$ in a given period:

$$\text{IC}_{\text{Pearson}} = \text{corr}(s_i, r_i) = \frac{\sum_i (s_i - \bar{s})(r_i - \bar{r})}{\sqrt{\sum_i (s_i-\bar{s})^2}\sqrt{\sum_i (r_i-\bar{r})^2}}.$$

It measures a linear relationship between raw signal and raw return, and is sensitive to any single extreme value, since squared deviations from the mean can be dominated by one large term.

Rank IC (Spearman) first converts both signal and returns to within-period ranks — $s_i \to \text{rank}(s_i)$, $r_i \to \text{rank}(r_i)$ — then computes Pearson correlation on those ranks. Because ranks are just integers $1, \ldots, n$ regardless of how extreme the underlying values were, a wildly outsized return (say, from a merger) contributes exactly the same "pull" as any other stock — it can only ever be ranked 1st or last, never "infinitely far" from the rest.

Worked example: one outlier changing the picture

Suppose a five-stock cross-section has signal values $s = (1, 2, 3, 4, 5)$ and returns $r = (0.5\%, 1.0\%, 1.5\%, 2.0\%, 40\%)$ — the last stock had a huge spike from an acquisition, unrelated to signal quality. Pearson IC on these raw numbers is dominated by that one 40% return, computing to roughly 0.75 — driven almost entirely by that one pairing, even though the ordering among the first four stocks is already a clean, perfect match. Rank IC instead converts returns to ranks $(1, 2, 3, 4, 5)$ — the 40% return is simply "rank 5," no different in weight from a 3% return — and since signal ranks are also exactly $(1,2,3,4,5)$, Rank IC comes out a perfect 1.0, correctly reflecting that the signal got the ordering completely right regardless of the one extreme magnitude.

What this means in practice

Most systematic equity strategies report Rank IC as the primary signal-quality metric, since portfolio construction cares about relative ordering rather than exact magnitudes, and merger, earnings-surprise, or delisting outliers appear in nearly every real cross-section. Pearson IC remains a useful diagnostic when a signal specifically needs to predict return magnitude — feeding a mean-variance optimizer, say — and a large gap between the two metrics is itself informative: a sign that a handful of extreme observations are driving the Pearson number.