Case · Finance
Credit scoring — auto-attached adverse action reasons
// PERSONA
Cho, Risk Modeling Lead — commercial bank
// INDUSTRY
Finance / Banking
// DATASET
case_05 · credit_scoring_log.csv · 96,400 rows
PROBLEM
The Problem
From 2026, Korea's FSS requires "adverse action explainability" at the model layer for AI-based credit scoring — not as a post-hoc patch. Existing boosted models were black boxes; if a customer disputed a decline, the entire model had to be retired.
COMING SOON
APPROACH
XimTier Approach
Credit, transaction, and income data are integrated in an air-gapped environment. SHAP-based per-variable contributions are auto-attached to every decision. Reverse What-If even generates customer-facing guidance ("which variable, by how much, to become approvable"), and regulator-submission documents are produced automatically.
COMING SOON
OUTCOMES
Outcomes
OUT / 01
Auto-generated adverse action notices
OUT / 02
Dispute resolution time −62%
OUT / 03
Fraud false-positive rate −38%
OUT / 04
Air-gapped operation (zero external LLM calls)
⚠ Example data — fine-tuned on customer data at deployment.