Information Coefficient (IC) and Rank IC for Quant Signals

For each date t, an analyst compares values of a signal across the eligible assets with returns those same assets earn over a later, defined horizon. A common convention is IC_t = corr_i(signal_i,t, forward_return_i,t+1). The subscript i indicates that the correlation is calculated across assets for one date, then dates form a time series of IC observations

This is a forecast-evaluation statistic for a cross section, not the correlation of two pooled time series and not an event-study abnormal return. Its value depends on the universe, signal definition, return horizon, weighting, and date-by-date coverage. State whether IC means Pearson correlation in your study; terminology is not perfectly uniform

Grinold’s original relationship is expected IR ≈ IC × √B under its assumptions, where B represents effective breadth—the number of sufficiently independent decisions (Grinold, “The Fundamental Law of Active Management”). A later generalized form adds a transfer coefficient, IR ≈ IC × TC × √B, to reflect how portfolio constraints affect the passage from forecasts to active positions (Clarke, de Silva, and Thorley, “Portfolio Constraints and the Fundamental Law of Active Management”). These model relationships are frameworks, not promises that a measured IC becomes investment performance

At each rebalance cutoff, freeze the signal using only information that was actually available by that timestamp. Respect publication times, vendor delays, revisions, and the exchange time zone. A value labeled with date t may not have been observable at the close of t; using its later revision or same-period return can introduce look-ahead

Pair signal_i,t with a return that begins after the signal could have been acted on. Specify the decision time, first executable price, return start and end, and whether the target includes dividends or other distributions. If the signal uses the closing auction, a return beginning at that same close may be unattainable without an explicit execution assumption

Rebuild the eligible universe as it was known on each date, including delisted assets when appropriate. Applying today’s surviving constituents to old dates, or using future index membership, can make a weak signal look stronger. Point-in-time alignment is part of the statistic’s definition, not a cleanup step after calculating IC

Pearson IC is the ordinary correlation between the cross-sectional signal values and later returns on each date. It captures linear co-movement and preserves information about signal magnitude, but a few extreme assets can strongly affect it. A nonlinear monotone relationship may still have a modest Pearson value

Rank IC first converts the signal and forward returns into within-date ranks, then computes their Pearson correlation. This is Spearman’s rank association (Spearman, 1904). It tests whether higher signal values tend to precede higher returns, with less sensitivity to the original scales and extreme magnitudes; tied values require an explicit average-rank or other tie convention

The two measures answer related but different questions. Pearson IC may be more relevant when signal scale matters for sizing, while Rank IC focuses on cross-sectional order. Neither automatically accounts for portfolio constraints, nonlinear transformations, sector neutrality, or how forecasts are turned into positions

All inputs in this example are hypothetical. At one date, suppose five assets have raw signal scores exactly equal to [1,2,3,4,5]—so scores and ranks happen to coincide—and next-period returns [1%,2%,3%,4%,-5%]. Rank 1 means lowest and rank 5 highest; return ranks are assigned in ascending order, so the return ranks are [2,3,4,5,1]

For Pearson IC, the signal mean is 3 and the return mean is 1%. Centered signals are [-2,-1,0,1,2]; centered returns, measured in percentage points, are [0,1,2,3,-6]. Their cross-product sum is -10, while their sums of squared deviations are 10 and 50. Therefore Pearson IC = -10 / sqrt(10 × 50) ≈ -0.447

For Rank IC, correlate signal ranks [1,2,3,4,5] with return ranks [2,3,4,5,1]. The rank deviations have zero cross-product sum, so Spearman Rank IC = 0. In this hypothetical cross section, the linear association is negative because the fifth asset’s large loss dominates magnitudes, while the rank association is zero because the ordering is neither consistently positive nor negative. If the raw scores instead were [1,2,3,4,100], the ordering—and therefore Rank IC—would stay the same, but Pearson IC would change because the signal magnitudes changed

Calculate one cross-sectional IC_t per date, then summarize its time series. The arithmetic mean IC is mean_t(IC_t); also report the median, dispersion, fraction of positive dates, and a time plot when useful. A mean can conceal long weak periods or a short interval that accounts for most of the apparent skill

For a concrete convention, this guide uses weekly IC observations at a prespecified rebalance schedule and fixed forward-return horizon. Give every valid weekly date equal weight, then divide the arithmetic mean by the sample standard deviation (variance denominator n−1): ICIR_weekly = mean(IC_t) / s(IC_t). The ratio is unannualized. Report the schedule, return horizon, evaluation period, and fixed missing-date rule. Other weighting, cadence, or annualization conventions should be labeled; serial dependence can make square-root scaling unreliable

ICIR describes the consistency of date-level IC, not a portfolio Sharpe ratio or portfolio information ratio. Signal quality can vary over time, and the mean alone may not capture that variation; Ye studies how variation in signal quality affects performance (Ye, 2008). Ding and Martin’s revised treatment also makes time-varying IC central to its model (Ding and Martin, 2017)

At each date, match signals and forward returns for the same asset and remove only observations that fail a predeclared availability rule. Do not replace an unavailable future return with zero. Report the number of assets per date, coverage changes, treatment of stale prices and delistings, and how tied ranks are formed; changing cross-sectional composition can change the meaning and precision of IC

When next-period returns span multiple days, adjacent dates may have overlapping target intervals. The same asset can then contribute to several IC observations whose returns share price changes, creating serial dependence. Repeated assets and common date shocks also make observations dependent, so a naive standard error based on independent dates can overstate precision

For example, with weekly rebalancing and a 20-trading-session forward return, neighboring weekly targets share about 15 daily returns. Fifty-two weekly IC observations therefore do not represent 52 independent experiments. Rebalancing exactly 20 trading sessions apart removes this target overlap, but leaves fewer dates and may not remove common-market shocks or repeated-asset dependence. Match the cadence to the decision being studied and reflect remaining dependence in uncertainty estimates

Choose the horizon and overlap policy before examining results. Where overlap is necessary, use inference that reflects the dependence, such as an appropriate block or time-series procedure, and avoid treating each asset-date pair as independent evidence. Purging overlapping labels between training and validation periods can help prevent leakage; see the purged cross-validation guide

A high in-sample IC can come from tuning the signal to noise in the same historical sample. Reserve genuinely untouched periods or use walk-forward evaluation: at each step, fit transformations and choose parameters only with earlier data, then evaluate on later dates. Do not repeatedly inspect a holdout and keep adjusting the signal while still calling it out of sample

Count the research choices that were tried: signal definitions, lags, horizons, universes, filters, neutralizations, and date ranges. Testing many candidates increases the chance that some display a high IC by luck. Harvey, Liu, and Zhu explain why conventional significance cutoffs can be misleading after extensive factor mining and multiple testing (Harvey, Liu, and Zhu, 2014)

Report the selection process, the number or scope of trials, and uncertainty adjusted for serial dependence and search where feasible. A positive mean IC in one backtest is evidence about that tested sample and design; it is not confirmation that the signal will persist in a new market regime

Report the point-in-time universe, signal timestamp, execution convention, return horizon, Pearson or Spearman choice, tie handling, cross-sectional sample size, date-level IC series, ICIR formula, missing-data policy, and out-of-sample protocol. Include uncertainty estimates that reflect date and asset dependence and disclose how many variants were tried

IC measures association between a forecast and subsequent returns; it does not specify position sizes. Portfolio construction maps forecasts into weights, and results then depend on risk controls, neutrality constraints, turnover, transaction costs, liquidity, capacity, borrow and execution. A useful Rank IC may be unprofitable after costs, while a low average IC may still combine with breadth and portfolio design in non-obvious ways

Keep this evaluation distinct from a single-event CAR. An event study measures abnormal returns around a defined announcement; IC evaluates whether repeated cross-sectional forecasts order later returns. See the event-study CAR guide for that separate measurement. Likewise, IC is predictive association rather than proof of a causal mechanism; the correlation-versus-causation guide discusses that broader distinction