Real-Time Credit Risk Decisioning at Checkout for Short-Term Installment Products: Feature Availability Constraints, Adverse Action Explainability, and Portfolio Loss Monitoring
DOI:
https://doi.org/10.70917/ijcisim-2026-5720Keywords:
Credit decisioning, real-time underwriting, buy now pay later, installment credit, feature availability, feature store, latency budget, adverse action, reason codes, explainability, fair lending, portfolio monitoring, vintage analysis, model risk management, champion challengerAbstract
Extending credit at the moment of purchase compresses an underwriting decision that traditionally takes minutes or days into a window measured in milliseconds, and every part of the decision has to change to fit. This paper sets out an architecture for checkout-time credit decisioning on short-term installment products, organised around three constraints that shape it. The first is feature availability: a highly predictive signal is worth nothing at checkout if it cannot be obtained, validated, and consumed inside the time budget, so availability, freshness, and latency must be treated as first-class properties of a feature rather than as engineering details discovered late. The second is adverse action explainability, which is a legal obligation rather than a product preference; the paper argues that decline reasons must be generated programmatically from the same policy and model outputs that produced the decision, through a pre-approved mapping, so that the explanation is deterministic, reproducible, and derived from the decision rather than reconstructed after it. The third is portfolio loss monitoring, where the governing view is cohort-based by policy version, because aggregate portfolio health can remain stable while a specific approval cohort deteriorates. The work draws on checkout-time decisioning across consumer payments products between February and July 2022, covering decisioning requirements, the data needed at decision time, integration with decisioning systems, and the latency envelope. No approval rates, latency percentiles, delinquency rates, loss curves, or policy-change impacts are reported, because measurements sufficient to substantiate them were not retained. The paper is therefore presented as a decisioning architecture with a stated evaluation design, and Section 12 defines the measurements that would establish it.
