Traditional weight of evidence assigns a credit characteristic a value based on its association with default across the whole population. The paper proposes a conditional version that instead measures the change in default log odds for a specific applicant after the rest of their credit record is already known. The sign of this conditional weight of evidence shows whether the characteristic adds adverse or favorable information for that applicant, and the magnitude indicates how strong that information is. The same characteristic value can therefore receive different assessments for applicants with different payment histories or other credit attributes.
The authors develop estimators using gradient-boosted trees and evaluate them in a bureau-style simulation. They find that regressions of the full model's predicted probabilities recover conditional evidence more accurately than separate models fitted directly to default outcomes, provided the full model is accurate. This approach also recovered a deliberately constructed file-age effect whose direction varies with utilization and is often missed by marginal weight of evidence. Local class-frequency estimates improved on marginal weight of evidence but remained inaccurate for some weak effects.
An application to FICO HELOC data illustrates how the method works when population evidence is unknown. The paper also introduces a population summary, conditional information value, which splits the joint information in a record across blocks when averaged over orderings. The authors stress that neither conditional weight of evidence nor conditional information value is an adverse action reason; adverse action reasons explain the fitted score difference between the applicant and chosen passing records.