What Is Information coefficient (IC)?
The information coefficient (IC) measures the predictive skill of a forecast or signal. In quantitative finance, it is the correlation between a signal available at a decision time and the return subsequently realised over a specified holding period. An IC of zero indicates no linear or monotonic association; a positive IC means higher signal values tend to be followed by higher returns, while a negative IC means the reverse. IC values lie between -1 and 1, though useful financial signals usually have much smaller values.
The conventional IC is Pearson correlation between a model’s predictions and realised returns. A Rank IC instead uses the correlation between their ranks (commonly Spearman rank correlation). Rank IC is often preferred for cross-sectional stock selection because it tests whether the model orders opportunities correctly, without requiring its predicted return magnitudes to be perfectly calibrated.
In machine learning, IC is an out-of-sample evaluation metric for a model that predicts returns, relative returns, or asset scores. It complements loss measures such as mean squared error: a model can be useful for ranking assets even when its return estimates are not accurately calibrated. Researchers calculate IC on validation and test periods, inspect its stability across time and market regimes, and may use its mean and variability to compare models. The information-coefficient information ratio (ICIR) is commonly defined as mean IC divided by the standard deviation of IC across evaluation periods.
In algorithmic trading and a backtest, a cross-sectional IC is typically calculated at each rebalance date by correlating the signal values across the tradable universe with forward returns over the intended holding period. A persistently positive out-of-sample IC supports a long-high-signal, short-low-signal or overweight-high-signal portfolio rule; a negative IC may support reversing the signal. IC alone is not a trading result: a strategy must also be tested for turnover, transaction costs, capacity, risk exposure, and robustness. IC measurements must use only information available at each decision time and genuinely out-of-sample returns to avoid look-ahead bias and overfitting.
See also
Research and literature
Richard C. Grinold, The Fundamental Law of Active Management, The Journal of Portfolio Management 15(3), 1989.