IC Standard Errors and Significance

A researcher tells you a signal has an information coefficient of 0.04 over the last five years. Is that a real edge or a number that fell out of noise? The instinct is to treat the 0.04 as settled fact, but an IC is a correlation computed from a finite, messy sample, and every correlation you have ever met has an error bar attached. The question is not "what is the IC" but "how sure are we".

What an IC actually is

The information coefficient in month $t$ is the cross-sectional correlation between a forecast made at the start of the month and the return that followed it: line up every stock's predicted rank against its realised rank and see how well they agree. That single number, $IC_t$, already compresses an entire month's cross-section into one statistic — it says nothing about how many stocks went into it. Do this every month for $T$ months and you have a short time series, $IC_1, \dots, IC_T$, and it is that series whose properties you need to test, not the underlying stock-months.

This is the same trap the Fama-MacBeth Regression page warns about, wearing different clothes. A signal-research deck that reports "we tested this signal on 500 stocks a month for 60 months, so $n = 30{,}000$" has confused rows in a spreadsheet with independent pieces of evidence. Within a given month, the 500 stocks share the same market regime, so their contribution to that month's IC is already one collective vote, not 500 separate ones.

The two-line calculation

Treat the monthly IC series exactly like any other time series of returns. Its average is the estimated signal strength and its spread tells you how confident to be in that average:

$$\overline{IC} = \frac{1}{T}\sum_{t=1}^{T} IC_t, \qquad \text{SE}\big(\overline{IC}\big) = \frac{s_{IC}}{\sqrt{T}}, \qquad t = \frac{\overline{IC}}{\text{SE}\big(\overline{IC}\big)},$$

where $s_{IC}$ is the sample standard deviation of the $T$ monthly values. In plain English: the average IC is just the mean of the monthly numbers, and the uncertainty around it shrinks with the square root of the number of months, not the number of stock-months. A signal tested on 5,000 names a month for two years is worth less, statistically, than one tested on 200 names a month for ten years, because the second has more independent draws of the market regime.

Worked example

A signal is scored against forward one-month returns across roughly 500 stocks a month, for five years ($T = 60$ months). The monthly IC values average $\overline{IC} = 0.04$ with a sample standard deviation $s_{IC} = 0.09$.

Step 1. $\text{SE} = 0.09 / \sqrt{60} = 0.09 / 7.75 = 0.0116$.

Step 2. $t = 0.04 / 0.0116 = 3.45$. That comfortably clears 1.96, so the average IC is statistically distinguishable from zero.

Step 3 — the wrong way, for contrast. Someone pools all $500 \times 60 = 30{,}000$ stock-months and computes a Fisher-style correlation standard error, $\text{SE} \approx 1/\sqrt{30{,}000} = 0.0058$, giving $t = 0.04/0.0058 \approx 6.9$. That number looks better and is fabricated: it assumes 30,000 independent observations when there were 60. The honest calculation used the month-to-month scatter of the IC itself, which already carries the within-month correlation as its natural volatility.

In practice

• An IC of 0.03–0.05 can be genuinely significant if it is stable over many periods — it does not need to be large, it needs to be persistent.
• Report $t$-stats alongside the mean IC, computed from the period series, whenever comparing signals; a bare IC number invites exactly the pooling mistake above.
• The information ratio inherits this. $IR \approx \overline{IC}/s_{IC} \times \sqrt{\text{periods per year}}$ is the annualised version of the same $t$-statistic, which is why The Fundamental Law of Active Management and Estimating Breadth With Correlated Bets both build on this series, not on the stock count.
• Autocorrelation in the IC series itself (a slow-moving factor tilt, say) widens the true standard error further; see Effective Sample Size With Autocorrelated Observations.