The Information Coefficient answers one question. On the day you ranked the universe by a signal, did the ranking line up with what happened next?
It is computed cross-sectionally. On each rebalance date you take the signal value for every stock in the universe and the return each stock went on to deliver over the forward window, convert both to ranks, and correlate the ranks. That gives one IC per date. The reported figure is the average of those dates, usually with the standard deviation alongside it.
Ranks matter here. A single stock returning 300% would dominate a Pearson correlation and tell you nothing about the other 499 names. Rank correlation asks only whether the ordering held.
What the magnitudes mean
| Mean IC | Reading |
|---|---|
| 0.00 to 0.01 | No usable ordering |
| 0.02 to 0.05 | The realistic band for a genuine cross-sectional factor |
| 0.05 to 0.10 | Strong. Verify the data pipeline before believing it |
| Above 0.15 | Almost always a bug rather than a discovery |
Those numbers look disappointingly small, and they are supposed to. An IC of 0.03 means the signal explains a fraction of a percent of the cross-sectional variance in forward returns. It still compounds, because you apply it to 500 names every month rather than to one bet. A common approximation holds that the information ratio of a factor scales with the IC multiplied by the square root of the number of independent bets, so breadth is what turns a weak ordering into a portfolio result.
Why a high IC is usually bad news
An IC above 0.15 on an equity cross-section should send you to the data layer, not to the deploy button. The usual causes are look-ahead of some form. Fundamentals stamped with the period end rather than the filing date give you numbers months before the market had them. Restated financials silently overwrite what was originally reported. A universe built from today’s index constituents quietly drops the names that were delisted, so the signal appears to avoid disasters it never had to face.
Caveats
A single date’s IC is noise. For a universe of 500 names the standard error of a Spearman correlation is roughly 1 / sqrt(n - 1), about 0.045, which is larger than the signal you are hunting. Only the average across many dates carries information, and you need its t-statistic to say whether the average is real.
IC treats the whole ranking equally. A factor can post a respectable IC while the tradeable extremes, the top and bottom deciles you would actually hold, behave badly. Read it next to a quantile sort rather than on its own.