The Information Coefficient (IC)

The Information Coefficient (IC) is the workhorse statistic of cross-sectional factor analysis. It measures how well a factor score predicts subsequent realised returns — not in absolute terms, but in ranking. A factor with an IC of 0.05 averaged across many periods is genuinely useful; the same factor scoring 0.30 in a single period is almost certainly noise.

Definition

For a single rebalance period, IC is the Spearman rank correlation (Spearman 1904) between the cross-section of factor scores at the start of the period and the cross-section of realised returns over the period:

Spearman is preferred over Pearson because (1) factor scores are usually constructed by ranking or z-scoring within a universe and have heavy tails, (2) returns have heavy tails (Cont 2001), and (3) rank correlation is robust to the cross-sectional outliers that occasionally dominate Pearson correlation.

The Fundamental Law of Active Management

Grinold (1989) showed that the information ratio (IR — the active version of Sharpe) of an active portfolio is approximately:

where N is the number of independent forecasts per year. This is the most quoted relationship in active management. It says: a small consistent edge applied to many independent bets produces a large information ratio. An IC of 0.04 (small but persistent) applied to 250 daily forecasts produces an IR of 0.04 · 15.8 = 0.63 — competitive with most active managers.

The relationship has two implications often missed:

• Breadth matters as much as skill. A factor with IC = 0.08 applied quarterly (N=4) produces IR = 0.16. The same factor applied daily produces IR = 1.27 — if the daily applications are independent. They usually aren't.
• N is "independent" forecasts, not "total" forecasts. If you rebalance daily but holdings change slowly, your effective N is far lower than 250. This is why daily-rebalance factor strategies rarely deliver the IR the math suggests.

What's a "good" IC?

Practitioner norms (Grinold & Kahn 1999):

• IC ~ 0.02: Marginal, likely indistinguishable from noise at typical sample sizes.
• IC ~ 0.04–0.06: Real signal. Most academic factors fall here.
• IC ~ 0.08–0.12: Strong signal. Rare but achievable with multi-factor composites or alternative data.
• IC > 0.15: Suspicious. Either a genuinely powerful undiscovered signal (very rare) or a methodology bug (look-ahead, survivorship, in-sample leakage).

IC statistical significance

With N rebalance periods and IC standard deviation σ_IC, the t-statistic for "mean IC is non-zero" is:

For a typical 20-period backtest with mean IC = 0.05 and per-period IC standard deviation of 0.10, t = 0.05 · sqrt(20) / 0.10 = 2.24 — just over the 2-sigma threshold. A 5-period backtest with the same numbers has t = 1.12, well within noise. Always look at the t-statistic alongside the IC.

IC time-series properties

The Factor Diagnostics sub-pill plots IC by period. Three patterns to look for:

1. Stable mean, low variance. Healthy. The factor works consistently.
2. Drifting mean. Factor is decaying (or, in the bad case, the factor used to work and stopped). See Factor Decay.
3. High variance with regime cluster. Factor works in some regimes but not others. Cross-reference with Regime-Conditional Performance.

Common IC mistakes

• Reporting a single IC number. A single IC is a single observation. Always report mean, t-statistic, and time-series plot.
• Computing IC with overlapping periods. If holding period is longer than rebalance frequency, IC observations are not independent — the t-stat overstates significance. Use Newey-West (1987) standard errors for overlapping observations.
• Equally weighting IC across regimes. Aggregate IC averages across periods regardless of how unusual each period was. Better: compute IC within macro regimes (high/low vol, expansion/recession) and look at the distribution.
• Confusing IC with hit rate. Hit rate (fraction of periods where the top quintile beat the bottom quintile) is a related but different statistic. A factor can have IC = 0.05 with hit rate = 55%, or hit rate = 80% with much higher per-hit return.

Top-vs-bottom quintile spread

An alternative to Spearman IC is the spread between the top and bottom quintile returns. The "quintile spread" is easier to communicate (it's a return number, not a correlation) but discards the middle of the distribution. The IC and the quintile spread usually move together — if they don't, that's a signal the factor is acting non-monotonically (e.g., extreme high and extreme low both win for different reasons).