Factors, IC and quantile evaluation

A factor expresses a security characteristic as a number. Price momentum uses past price changes as a feature. Calculate features from information available at the decision time and keep them separate from labels describing later returns. Using future labels as features introduces knowledge of the outcome.

Daily cross-sectional IC measures correlation between factor values and forward-return labels across securities on one date. Rank IC compares their ranks. Interpretation depends on factor direction and label definition; correlation alone does not establish trading returns.

Quantile evaluation groups securities by rank and checks returns, sample counts, missingness and changing membership. Three securities permit at most three quantiles. Code-order tie breaking creates no predictive information, and overlapping return windows do not create independent observations.

If three securities have factor values 1, 2 and 3 and later returns −3%, −2% and −1%, their ranks agree and Rank IC is 1. Yet every security loses value. Rank association and absolute returns are different; correlation is undefined when a series is constant.

Evaluation labels

A label describes the later outcome being evaluated. For a five-session return, specify its starting and ending dates and prices. Forward returns are evaluation outcomes, not decision-time features.

Cross-section and coverage

A cross-section contains several securities on the same date. Coverage is the fraction of the target set with valid values. Eight valid securities out of ten is 80%; inspect the universe and validity rules.

ICIR

This course uses mean daily IC divided by its standard deviation as a descriptive statistic. A zero denominator makes it undefined. Annualization and statistical inference require additional assumptions; overlapping labels are not independent samples.

Read the screenshot-led tutorial · All basic concepts · Research methodology

Correlation definitions: Pearson correlation · Spearman rank correlation