Do not apply merchant-only features to an individual credit applicant
Alternative-data features need an explicit population and missingness contract before their absence can be interpreted.
Access promises depend on input meaning
Using alternative evidence for people with limited financial history requires more than adding new signals. The team must establish to whom each feature applies. An individual’s lack of merchant-only activity is not evidence of poor credit. Truthound can check the declared population and missingness contract in an authorized feature extract. It does not score an applicant, establish lending eligibility or determine whether a model is fair.
Designation and expected benefits are different
Korea’s Financial Services Commission announced a designated-agent approval for an alternative-data credit-assessment service on 27 August 2026. The release distinguishes merchant store activity from individual payment or content activity, with financial institutions retaining lending decisions. Improved financial access is an expected benefit, not an outcome established by the approval itself. The data contract should preserve those differences before anyone interprets missing features.
Put applicability into the feature contract
If authorized staff prepare an extract, a proposed table retains applicant type, feature type, observation window, source-permission reference and missing reason. Separate an empty value from a feature that does not apply. The same person could also operate a business, but the role relevant to the particular application still matters. A global person identifier alone should not silently determine which feature population the application belongs to.
Examine a consequential counterexample
If an individual application lacks store visitor counts, classifying that absence as collection failure may create an unnecessary request for information. Conversely, calling a genuinely required merchant feature inapplicable can hide an incomplete input. Responsible staff must define the evidence that distinguishes those states. These are counterexamples for contract design, not findings about an applicant’s creditworthiness or a verified defect in the designated service’s operation.
Separate a passing check from a lending decision
Truthound Core Python SDK 3.3.4 (open-source SDK, Polars backend) documents ColumnPairInSetValidator for approved applicant-feature pairs. CompletenessRatioValidator requires a caller-defined minimum non-null ratio appropriate to the task and population. Do not apply one global completeness target as if inapplicable features were failures. A valid category or format does not establish that the evidence is true, that a prediction is accurate or that its consequences are fair.
Consider the feature validation already maintained
Pandera 0.35.0 (MIT open-source Python DataFrame validation library; pandas, Polars and other supported backends) custom checks can express the same applicability and missing-reason rules. Where model and data owners already maintain that contract, retain it and identify only the control still missing. Compare Truthound at the feature-admission layer. Neither documentation set supports a claim that its validator improves lending access, delivers a more accurate credit model or makes consequential decisions fairer for the affected people.
Clarify absence before assigning consequences
Start with one feature for one application type. Distinguish inapplicable, uncollected and clarification-required states, and have the owner describe a suitable correction route. Where evidence remains ambiguous, hold the input interpretation for specialist review. The lasting improvement is a structure that keeps asking whose circumstances a feature describes and what it may mean, rather than assuming that collecting more data supplies those answers automatically.
Conditions
Define applicant-feature scope, authorized sources and missing reasons before evaluating an extract.
Alternatives
Pandera 0.35.0 custom checks can retain the same applicant-feature applicability contract.
Limitations
Missingness or format conformance establishes neither creditworthiness, fairness nor lending eligibility.
Evidence
- Event checked
- Subject checked
- Modified
