Guide

Information coefficient explained

Harbor Capital's systematic equity team pitched a value-momentum composite in Q1 2026. The backtest showed 14% annualized alpha and a Sharpe above 2.0 after costs — numbers that cleared the allocator's headline hurdles. Then risk asked a sharper question: did the signal actually rank stocks correctly each month, or did a few lucky quarters carry the backtest? Portfolio-level Sharpe and information ratio measure realized fund performance relative to a benchmark over time. The information coefficient (IC) measures something earlier in the pipeline: how well a cross-sectional forecast correlates with subsequent realized returns at each rebalance date. Stable positive IC is the raw material of scalable quant alpha; unstable or negative IC periods explain why many “great” backtests fail live. This guide defines IC and its stability ratio ICIR, contrasts Spearman and Pearson estimation, links IC to Grinold-Kahn breadth and turnover, works a Harbor Capital factor review, provides a metric decision table, lists pitfalls, and ends with a quant diligence checklist. For portfolio construction from factors, see factor investing; for backtest hygiene, see backtesting trading strategies.

What the information coefficient measures

At each rebalance date t, you have a vector of forecast scores f_i,t for stocks i = 1 … N and a vector of realized forward returns r_i,t+1 over the holding horizon (often one month). The information coefficient is the correlation between forecasts and outcomes across stocks at that date:

IC_t = corr(f_1,t, …, f_N,t ; r_1,t+1, …, r_N,t+1)

A positive IC means higher-scored names tended to outperform lower-scored names over the forward window. IC near zero means the signal had no cross-sectional ranking power that period. IC is typically small in magnitude — even excellent institutional signals often average 0.03 to 0.08 monthly — because stock returns are noisy and the measurable edge per name is thin.

Spearman vs Pearson IC

Spearman rank IC (the industry default) correlates rank-transformed forecasts with rank-transformed returns. It is robust to outliers, fat tails, and nonlinear monotonic relationships — a forecast that correctly orders deciles but mis-scales magnitudes still scores well. Pearson IC correlates raw values and rewards linear calibration; use it when forecast scores are intended to map directly to expected return percentages. Document which you report; mixing them across vendors invalidates comparison.

IC vs information ratio

These names confuse allocators because both use “information.” IC is cross-sectional (across stocks at one date). The information ratio is time-series (active return of a portfolio vs tracking error over many periods). High average IC with poor implementation — excessive turnover, capacity limits, crowding — can still produce a low IR. IC diagnoses the signal; IR diagnoses the fund.

ICIR: stability of the signal

A single month's IC is nearly meaningless. Quants summarize a signal's history with two statistics:

Mean IC: average IC_t across T rebalance dates.
IC volatility: standard deviation of IC_t across dates.
ICIR (information coefficient information ratio): mean IC divided by IC volatility, sometimes annualized by multiplying by √12 for monthly rebalances.

ICIR answers: is the signal consistently right, or did a handful of heroic months create a positive mean? Rule-of-thumb screens (not laws of physics) used in allocator diligence:

Mean IC > 0.02 monthly for single-factor equity signals warrants deeper review.
ICIR > 0.5 (monthly, unannualized) suggests reasonable stability; below 0.3 often indicates a fragile or regime-dependent signal.
Hit rate: fraction of months with IC > 0; strong signals often exceed 55–60% without being 80% (that may indicate overfitting).

Plot the cumulative IC (running sum of monthly IC) and a histogram of monthly IC values. A signal with positive mean IC but long stretches of negative cumulative IC is psychologically hard to run and may blow up in live trading when the PM capitulates during a drought.

IC decay, horizon and turnover

IC is not a single number for all time horizons. Compute IC at multiple forward windows — 1-day, 5-day, 21-day, 63-day — to map IC decay. Fast-decay signals (high 1-day IC, collapsing 21-day IC) require low-latency execution and tight turnover controls. Slow-decay signals tolerate monthly rebalancing but compete with capacity and crowding.

Link to Grinold-Kahn fundamentals

The Grinold-Kahn framework relates expected active return to:

IR ≈ IC × √BR

where BR (breadth) is the effective number of independent bets per year. Doubling universe size or rebalance frequency increases breadth but may reduce IC if the signal is diluted or costs rise. This identity explains why allocators push for both decent IC and broad, repeatable application — a 0.10 IC on 50 names annually may lose to a 0.04 IC on 2,000 names if implementation scales.

Turnover and transaction costs

High IC at the signal level does not survive 300% annual turnover if spreads and market impact consume 200 bps. Pair IC analysis with simulated portfolio turnover and net alpha after realistic cost assumptions. Signals with similar mean IC but different score autocorrelation can have vastly different live economics.

Harbor Capital worked example

Harbor's allocator reviewed the proposed value-momentum composite on 1,847 U.S. mid-cap names, monthly rebalance, 1998–2025 history, Spearman IC with 21-trading-day forward returns, universe filtered for liquidity (>$5M ADV).

Mean monthly IC: 0.041 (value leg 0.028, momentum leg 0.035, composite 0.041 after z-score blend).
IC volatility: 0.078 → ICIR = 0.53 (0.53 × √12 ≈ 1.84 annualized).
Hit rate: 58% of months IC > 0.
Worst 12-month cumulative IC: −0.31 (2018 Q4 value crash); composite recovered within 14 months.
Simulated long-short decile spread: 9.2% annualized gross, 4.1% net after 85 bps round-trip cost assumption at 180% turnover.

The backtest Sharpe of 2.0 was credible once IC stability was confirmed, but the 2018 drought flagged regime risk: value IC went negative for nine consecutive months while momentum IC stayed positive — motivating a dynamic weight cap rather than static 50/50 blending. Harbor approved a capped live sleeve with half the backtest leverage and a hard stop if rolling 24-month ICIR fell below 0.25.

Metric decision table

Your question	Start here	Also check
Does this factor rank stocks before portfolio construction?	Mean IC and ICIR on out-of-sample window	Hit rate, cumulative IC plot, subperiod stability
Signal works in backtest but allocator doubts robustness	IC by year and regime (bull/bear, high/low vol)	Backtest pitfalls — lookahead, survivorship
Compare two quant managers with similar IR	Disclosed IC/ICIR if available; else ask for signal diagnostics	Turnover, capacity, crowding, cost model
High-frequency vs monthly factor	IC decay curve across horizons	Implementation shortfall, spread costs
Combine multiple alpha signals	IC of composite vs legs; correlation of monthly IC series	Orthogonalization, dynamic weighting rules
Fund-level performance vs benchmark	Information ratio	Tracking error, upside/downside capture
Long-short equity capacity	IC on liquid sub-universe vs full universe	Long-short equity borrow and short rebate

Common pitfalls

Lookahead bias. Using same-day returns or financials not yet public at forecast time inflates IC; align data timestamps to point-in-time databases.
Survivorship bias. Computing IC on today's index members retroactively removes delisted bankruptcies that often had extreme negative realized returns.
Microcap illusion. High IC on illiquid tails vanishes after ADV filters; always report IC on the tradable universe.
Overlapping return windows. Non-independent monthly IC observations from overlapping 21-day returns understate IC volatility; use non-overlapping samples or Newey-West adjustments for ICIR.
Outlier sensitivity. One merger-arb name can dominate Pearson IC; prefer Spearman or winsorize returns at 1–99%.
Regime cherry-picking. Reporting IC only on 2009–2021 bull market hides 2008 and 2022 stress; require full-cycle disclosure.
Confusing IC with R-squared. IC of 0.05 does not mean the signal explains 0.25% of variance in a simple regression sense across a noisy panel; keep definitions straight in IC memos.
Ignoring implementation. IC without turnover and cost simulation is an academic exercise, not an allocator-grade diligence pack.

Quant diligence checklist

State forecast horizon, rebalance frequency, and universe rules (liquidity, ADR inclusion).
Report Spearman IC by default; note if Pearson is also shown.
Provide mean IC, IC volatility, ICIR, and monthly hit rate over at least 10 years.
Plot cumulative IC and histogram of monthly IC values.
Break IC by calendar year and by macro regime (e.g., VIX quartiles).
Map IC decay across multiple forward return windows.
Disclose point-in-time data vendor and corporate-action handling.
Run IC on liquid sub-universe to test capacity.
Pair IC analysis with simulated turnover and net spread costs.
Document signal combination weights and IC correlation between legs.
Set live monitoring triggers on rolling 24-month ICIR (Harbor uses 0.25 floor).

Key takeaways

Information coefficient measures cross-sectional ranking skill — how well forecast scores align with subsequent returns at each rebalance date.
ICIR (mean IC over IC volatility) separates durable signals from backtests carried by a few lucky months.
IC and information ratio answer different questions: IC grades the signal; IR grades the live portfolio relative to a benchmark.
IC decay, breadth, and turnover determine whether a statistically real signal survives costs at scale.
Allocator-grade IC analysis requires point-in-time data, full-cycle disclosure, and pairing signal metrics with implementation economics.