Informationswerte

Informationswerte, also known as Information Value (IV) in English, is a statistic used to measure the predictive strength of a feature in binary classification tasks. It is widely used in credit risk modeling and feature selection to assess how well a variable separates the two outcomes.

Computation and concept: For a continuous variable, data are typically binned into categories. For each bin, the proportions of the two outcomes (for example, good vs. bad, or non-defaulters vs. defaulters) are computed relative to their total counts. The Weight of Evidence (WOE) for bin b is defined as WOE_b = ln( Good_b / Bad_b ), where Good_b is the proportion of non-defaulters in bin b and Bad_b is the proportion of defaulters in bin b. The Information Value is IV = sum over all bins of ( Good_b − Bad_b ) × WOE_b. Some formulations use alternative notations, but the underlying idea is the same: IV aggregates how differently the two outcomes are distributed across the bins.

Interpretation and use: A higher IV indicates stronger discriminatory power of the variable. Common (though not universal) guidelines for IV interpretation are: IV < 0.02 not predictive, 0.02–0.1 weak, 0.1–0.3 medium, 0.3–0.5 strong, above 0.5 very strong. Thresholds can vary by domain and sample size. IV is often used for feature selection and to guide binning and Weight of Evidence transformations in logistic regression and related models.

Caveats: IV depends on the binning method and sample size and does not directly measure overall model performance. It can be affected by data leakage or drift. It is best interpreted alongside other metrics such as ROC-AUC, KS, and calibration.