What Is IC? Reading the Relationship Between Factors and Future Returns

IC, or information coefficient, measures correlation between factor values and subsequent returns. In the cross-sectional setting here, it asks whether higher-scored securities on the same date subsequently had higher returns.

Understand the score first: a factor value, rank and portfolio rule are different concepts.

Here daily IC is Pearson correlation between security-level factor values and future-return labels on one date. A cross-section compares different securities at the same time, unlike a time series that compares one security across dates.

Pearson correlation describes a linear relationship and, when defined, lies between −1 and +1. Positive values associate higher scores with higher subsequent returns; negative values indicate the opposite. Near zero means a weak linear relationship in that sample, not proof that every kind of relationship is absent.

Inspect daily variation, valid-day counts, security counts and missingness alongside the period mean. See factor and IC basics.

Rank IC is Spearman correlation between factor values and future returns. It examines ordering rather than numerical spacing. Ties require a ranking convention; conventional Spearman correlation uses average ranks.

Consider fictional securities EX-A, EX-B and EX-C with factor values 30, 20 and 10 and future returns +5%, +1% and −2%, respectively. They share one decision date and return horizon. Both rankings are A, B, C, so Rank IC is +1.

The factor values are spaced by 10, while return differences are four and three percentage points. Matching ranks do not necessarily create an exact linear relationship, so Pearson IC need not equal Rank IC. This invented example illustrates ranking, not real predictive performance.

A negative IC does not automatically mean a bad factor. If the design prefers lower raw values, a negative correlation can match that direction. Reversing the raw value into a higher-is-preferred score also reverses the correlation sign.

The return horizon and start and end prices can change the relationship. Specify direction and label definitions before evaluation. Selecting direction after viewing results fits the observed period.

The second public KabuForge course calculates features through the closing price of the session before the signal date, or D-1. Evaluation labels measure simple returns from execution open to the open of the fifth subsequent observed session. Features and labels are separate files; labels are used only in evaluation.

Feeding future outcomes into selection on the same date uses information unavailable at the decision time. Rows lacking future prices at the end stay missing. See data availability and look-ahead bias.

Correlation is undefined if all factor values are constant, or all future returns are constant. Rows with missing features or labels are unavailable for the comparison. Check the reasons when too few valid pairs or other conditions prevent evaluation.

Replacing an invalid day with zero changes the mean and variation. Inspect reasons, valid security counts and coverage to understand what could actually be compared.

Positive IC does not guarantee that the held securities rose. If all securities fell, higher-scored securities falling less could still create positive correlation. IC describes a relative relationship, not portfolio return.

Portfolio weights, rebalancing, fill conditions, commissions and slippage affect the trading result. Evaluate holding rules and costs and turnover separately.

With three securities, changes in one value or rank can strongly affect the result. High IC in this teaching sample does not establish stability across a broad universe or another period.

Daily five-session forward labels overlap. They share price movements, so day counts cannot be treated as independent sample counts. In this course, ICIR—mean daily IC divided by its standard deviation—is descriptive; it does not by itself establish statistical significance.

Repeatedly selecting windows, direction or securities from the results can capture historical chance. Fix definitions and inspect other periods, as discussed in overfitting and out-of-sample evaluation. Be able to explain which feature, information date and future-return interval were compared.

The public example comes from the second KabuForge GUI research course. The fictitious ranking example was created for explanation. Correlation definitions: SciPy Pearson correlation and SciPy Spearman correlation.

This article teaches research procedures, does not recommend trades and does not guarantee future returns. Disclaimer.

Explore the factors learning path

• What Is Quantile Analysis? Evaluating Factors by Groups
• Factor Coverage: Check Which Securities and Dates Reach Evaluation
• Rank Normalization: Comparable Scales Still Need Clear Conventions