Quant Analytics: from scratch to pro · Indicators, honestly measured · Pro
Information coefficient, breadth and signal decay
The information coefficient measures how well a signal ranks assets by their later returns. The fundamental law of active management links it to performance through breadth, the number of independent bets, and signal decay says how fast the information goes stale.
In this lesson
- Compute a rank information coefficient from signal ranks and return ranks
- Apply IR = IC x sqrt(breadth) and state the assumptions behind it
- Estimate the uncertainty of a mean IC across periods
- Model signal decay with a half-life and read its consequence for holding period and turnover
Read first: Does an indicator carry information? Measuring it honestly, Regression, beta and alpha, How sure is an average? Standard error and confidence intervals
Most quantitative signals are not about one asset. They rank many assets by what is expected to do relatively better. For a ranking signal, the useful questions are how good the ranking is, how many independent times a year it can be used, and how long the information lasts. The information coefficient, breadth and decay answer those.
The information coefficient
The information coefficient (IC) is the correlation, across assets at one date, between the signal and the returns that follow. A rank IC uses ranks instead of raw values (the Spearman correlation), which keeps one extreme value from dominating. An IC of 1 is a perfect ranking, 0 is no information and -1 is a perfectly wrong one.
Worked example
Six assets have signal ranks 1 to 6. Their later returns rank 2, 1, 4, 3, 6, 5. The squared rank differences are 1 each, summing to 6. The rank IC is 1 - 6 x 6 / (6 x 35) = 0.829. That is an unusually good ranking, chosen to make the arithmetic clear.
Size of an IC
Real ICs are small. An IC of 0.05 explains 0.05^2 = 0.25% of the variance of cross-sectional returns. A signal with a small IC can still matter, because it can be applied again and again. The IC for each date is then averaged over many dates.
Breadth and the fundamental law
Grinold (1989) showed how skill and the number of independent bets combine. In its simplest form the fundamental law of active management says the expected information ratio is
IR = IC x sqrt(breadth)
where breadth is the number of independent decisions per year. An IC of 0.05 with 100 independent bets gives an IR of 0.50. With 400 bets it is 1.00. An IC of 0.02 on 2,500 bets also gives 1.00. A small edge applied widely can equal a larger one applied rarely; the chart shows the square-root payoff.
The law rests on assumptions, and the result is only as good as they are:
- Independent bets. Breadth counts independent decisions, not the number of assets. A hundred stocks that move together are far fewer than a hundred bets. Bets on a signal repeated every day on slowly changing information are not independent either.
- Equal skill. The same IC for every bet.
- No constraints and no costs. Clarke, de Silva and Thorley (2002) add a transfer coefficient for the loss from constraints. Costs lower it further.
- A known IC. In practice the IC is an estimate.
Treat the law as a way to see the trade-off, not as a prediction of performance. A realised ratio below the formula's value is normal.
Uncertainty of a mean IC
Compute the IC each period and average it. Suppose the mean is 0.05, from 60 periods, and the IC across periods has a standard deviation of 0.10. The standard error of the mean is 0.10 / sqrt(60) = 0.0129 and the t-statistic is 3.87. A mean IC needs the same care as any average: check overlap between periods, and the number of variants tried (see multiple testing).
Signal decay
Information goes stale. The IC of a signal against returns t days later usually shrinks as t grows. If it halves every h days (the half-life), then IC(t) = IC(0) x 0.5^(t / h). With IC(0) = 0.05 and h = 5, the IC after 10 days is 0.0125, and the signal keeps 87.1% of its strength from one day to the next.
The shape is only a model. Measure it directly: compute the IC of the same signal against returns at several horizons and plot it. Three things follow from the profile.
- Holding period. Holding much longer than the half-life means holding a position whose reason has weakened.
- Turnover. Fast decay forces fast trading to keep the signal fresh, and turnover is paid for in costs (see costs in a backtest).
- Capacity. The same cost limits the amount that can be traded.
A signal with a small IC and slow decay can beat one with a larger IC and fast decay, once costs are included.
Common mistakes
- Counting assets as breadth when the bets are correlated.
- Quoting the IR from the law as an expected result.
- Using the IC from one period as the skill of the signal.
- Choosing the horizon of the best IC after looking at all horizons.
- Ignoring costs when the half-life is short.
In practice
For a ranking signal, report the mean rank IC with its standard error, the IC at several horizons, the number of independent periods and the number of variants tried. Use the law to reason about how many bets are needed, not to promise a ratio.
Information ratio from the fundamental law for an IC of 0.05
Signal decay with a half-life of 5 days
Worked numbers
Each result is recomputed from the formula shown.
| What | Formula | Result |
|---|---|---|
| sum of squared rank differences | (1-2)**2 + (2-1)**2 + (3-4)**2 + (4-3)**2 + (5-6)**2 + (6-5)**2 | 6 |
| Spearman rank correlation | 1 - 6 * d2 / (6 * (6**2 - 1)) | 0.829 |
| IR for IC 0.05 and breadth 100 | 0.05 * sqrt(100) | 0.50 |
| IR for IC 0.05 and breadth 400 | 0.05 * sqrt(400) | 1.00 |
| IR for IC 0.02 and breadth 2,500 | 0.02 * sqrt(2500) | 1.00 |
| variance explained by an IC of 0.05 | 0.05**2 | 0.25% |
| standard error of a mean IC over 60 periods | 0.10 / sqrt(60) | 0.0129 |
| t-statistic of a mean IC of 0.05 | 0.05 / se_ic | 3.87 |
| IC after 10 days with a half-life of 5 days | 0.05 * 0.5**(10 / 5) | 0.0125 |
| share of the signal kept after one day | 0.5**(1 / 5) | 87.1% |
Check yourself
Signal ranks of six assets are 1 to 6 and the ranks of their later returns are 2, 1, 4, 3, 6, 5. What is the rank IC?
The squared rank differences sum to 6. rho = 1 - 6 x 6 / (6 x 35) = 0.829.
A signal has an IC of 0.05 and 100 independent bets a year. What information ratio does the law give?
IR = 0.05 x sqrt(100) = 0.50. Quadrupling breadth to 400 doubles it to 1.00.
How much of the variance of returns does an IC of 0.05 explain?
IC squared = 0.0025, which is 0.25%. A useful signal can still be nearly invisible in a single period.
A mean IC of 0.05 comes from 60 periods whose IC has a standard deviation of 0.10. What is the t-statistic?
SE = 0.10 / sqrt(60) = 0.0129, so t = 3.87.
An IC of 0.05 at the time of the signal decays with a half-life of 5 days. What is it after 10 days?
0.05 x 0.5^(10 / 5) = 0.0125.
Key terms
Sources
- The Fundamental Law of Active Management. The Journal of Portfolio Management 15(3), 30-37 (Richard C. Grinold), 1989. Accessed 9 Nov 2026
- Active Portfolio Management (2nd edition). McGraw-Hill (Richard C. Grinold, Ronald N. Kahn), 2000. Accessed 9 Nov 2026
- Portfolio Constraints and the Fundamental Law of Active Management. Financial Analysts Journal 58(5), 48-66 (Roger Clarke, Harindra de Silva, Steven Thorley), 2002. Accessed 9 Nov 2026
- The Econometrics of Financial Markets. Princeton University Press (John Y. Campbell, Andrew W. Lo, A. Craig MacKinlay), 1997. Accessed 9 Nov 2026