A common first test of a stock characteristic is its rank information coefficient: each month, the Spearman correlation across stocks between the characteristic and next month's return. It is robust to outliers, which is why people use it. It also has a bias that is easy to miss.
What the data show
U.S. common stocks above the NYSE 20th size percentile, July 1963 to December 2020, from the factor research pipeline. Each factor is signed so that a higher value should mean a higher return.
| Factor | Rank IC (t) | Top-minus-bottom quintile, % per month (t) |
|---|---|---|
| Momentum | 0.032 (5.4) | 0.88 (4.4) |
| Volatility (low) | 0.028 (3.4) | 0.05 (0.2) |
| Beta (low) | 0.018 (2.3) | −0.04 (−0.2) |
| Size (small) | −0.010 (−2.1) | 0.22 (1.8) |
Momentum behaves as expected: both measures agree. The other three do not. Low volatility and low beta look like real signals by rank IC and earn nothing. Small size looks like a negative signal by rank IC and earns a positive spread.
Without the size cut the gap is wider. Across all stocks the rank IC of small size is −0.053 with a t-statistic of −10.0, while small stocks earn 0.71% a month more than large ones.
Why
A rank correlation asks, for a pair of stocks, which one did better. It does not ask by how much. So it measures typical outcomes, not average ones.
For a volatile stock those are far apart. Suppose the log return over a month is normal with mean and standard deviation . Then
Take two stocks with the same mean return, one quiet and one volatile. To have the same mean, the volatile one needs a lower , lower by half the difference in variance. Its median is therefore lower, and in a head-to-head comparison it loses more often than it wins. Its average is held up by a small number of months with very large gains, and a rank counts each of those as one win.
Any characteristic that sorts stocks from quiet to volatile inherits this. Low volatility and low beta sort that way directly. Size sorts the other way: small stocks are more volatile, so "small" gets a negative rank IC even though small stocks earn more on average.
What to do about it
- Report the spread next to the rank IC. The return of a top-minus-bottom portfolio is the number that corresponds to money. The pipeline's screen now prints rank IC, Pearson IC, and the quantile spread side by side.
- Be careful with anything weighted by rank IC. Weighting factors by their ICIR carries the tilt into the combined score. Methods fitted to returns, such as a regression, do not have it.
- Do not read this as "rank IC is wrong". For a factor that is unrelated to volatility, such as momentum here, it agrees with the spread and is less noisy.
This was found on in-sample data only, after the first screen showed size with the wrong sign. The check that settled it was the quintile returns: the data were fine, and the two measures were answering different questions.
Calculated based on data from CRSP US Stock Database, © Center for Research in Security Prices (CRSP), The University of Chicago Booth School of Business, and from Compustat, © S&P Global Market Intelligence.