Information coefficient: measuring the predictive accuracy of a systematic signal

A systematic strategy is only as good as the predictive accuracy of its underlying signal. The information coefficient measures that accuracy directly: the correlation between the signal's predicted ranking of assets and the assets' subsequent realised returns. It is the most fundamental diagnostic in quantitative strategy evaluation.

What the information coefficient is

The information coefficient (IC) is the correlation between predicted returns or signal scores and realised returns over a defined forward horizon. For a universe of N assets at time t, each asset has a signal score s_i,t and a realised return r_i,(t,t+h) over horizon h. The IC at time t is the cross-sectional correlation between the s_i,t and r_i,(t,t+h) across the N assets. Averaged over many time periods, the result is the strategy's average IC—the typical predictive accuracy of the signal.

The IC ranges between −1 and +1. An IC of zero means the signal has no predictive power; the rank ordering of asset returns is independent of the rank ordering of signal scores. An IC of +1 means perfect ordering; an IC of −1 means perfect reverse ordering (which is itself useful—invert the signal to extract the same predictive value). Realistic IC values for documented signals are very small in absolute terms.

Two variants are common: Pearson IC (using the actual values) and Spearman IC (using ranks). The Spearman version is more robust to outliers and is the more commonly reported figure in practical strategy evaluation.

How it works

For documented equity factors, IC values typically lie in the 0.02–0.05 range—that is, between 2% and 5% of the variation in cross-sectional returns is captured by the factor signal at the typical horizon. These figures sound modest but produce meaningful returns when applied across a large universe and over many periods.

Grinold's law makes the relationship explicit: information ratio ≈ IC × √breadth, where breadth is the number of independent bets per unit of time. A signal with an IC of 0.05 applied to 200 stocks each year produces an expected information ratio of 0.05 × √200 ≈ 0.71—a strategy that would deliver attractive risk-adjusted returns over time. The same IC applied to a small universe (10 stocks, breadth = 10) would produce an information ratio of 0.16—too small to be reliably distinguishable from noise. The same signal can be a profitable strategy or a useless one depending on how broadly it can be applied.

The corollary is that small IC improvements have outsized portfolio-level impact when breadth is large. A signal whose IC improves from 0.03 to 0.04—a 33% improvement that sounds dramatic in percentage terms—improves the information ratio by the same 33%, which over multi-year horizons compounds into meaningful additional return.

What the evidence shows

The momentum factor (Moskowitz, Ooi & Pedersen, 2012) has documented IC in the 0.04–0.07 range across the major asset classes the original paper studied. The value factor produces similar values in equity universes (Fama & French (1993) and subsequent work), and quality (Asness, Frazzini & Pedersen, 2019) and low-volatility (Frazzini & Pedersen, 2014) signals report similarly modest values. None of the documented systematic premia produce ICs above 0.10 on a regular basis.

The implication for retail investors evaluating strategies is direct. A strategy backtest claiming to predict returns with very high accuracy is almost certainly overfit. Real predictive signals operate at the 2–5% IC level; anything materially higher demands suspicion of in-sample data mining. The same logic applies to claimed Sharpe ratios: a backtested Sharpe of 3 or 4 in a standard liquid universe is a signal that the methodology has been calibrated against its own historical sample, not that the underlying edge is unusually strong.

IC is also unstable across time. A signal's average IC over a multi-decade evaluation window can hide substantial variation across sub-periods. The momentum factor, for instance, has produced IC near zero or even negative in some multi-year stretches (the 2009 momentum crash being the most prominent), even though the long-run average remains positive.

Limitations and trade-offs

IC is sensitive to sample size and to the choice of forward horizon. Computed on a small universe or over a short window, the figure has wide confidence intervals; even genuine signals can produce wildly varying IC estimates from period to period. Robust evaluation requires sufficient breadth and a long evaluation window—preferably out-of-sample.

The horizon choice matters. A signal can have positive IC at one horizon (say 1 month) and negative IC at another (say 6 months). The horizon at which a signal is informative determines the rebalancing frequency required to extract its predictive value. Rebalancing more frequently than the signal's horizon adds noise without adding signal; rebalancing less frequently captures only a fraction of the available information.

IC is also a measure of cross-sectional ranking accuracy, not of magnitude. A signal that correctly identifies which stocks will outperform but does not predict by how much will have positive IC; the same signal cannot tell the strategy designer how to size positions. Other diagnostics—predicted return magnitude, signal stability, transaction costs—complement IC rather than replace it.

Information coefficient in pfolio

Information coefficient is not currently displayed in pfolio Insights. The platform's systematic strategy evaluation focuses on portfolio-level Sharpe, drawdown, and CAGR rather than signal-level IC; users who want to evaluate the predictive accuracy of a specific signal can compute it externally from the asset's return series and the signal's historical values.

Related articles

• Factor investing explained: momentum, value, quality, and low volatility
• Backtesting investment strategies: methodology, limitations, and how to avoid overfitting
• Overfitting in quantitative investing: why backtested strategies fail in practice
• Time series momentum: the academic evidence behind trend-following strategies