Credit & Payments Risk Decision Engine

Underwriting Strategy · Fraud Monitoring · Expected Loss · Model Risk

A focused risk decision system that turns borrower and transaction data into underwriting decisions, fraud controls, expected-loss estimates, policy simulations, and model-validation evidence. It is a decision engine, not a prediction notebook — model scores are translated into governed actions under loss, fraud, calibration, drift, and review-capacity constraints.

• Live site: https://credits-engine.netlify.app
• Repository: https://github.com/VM25/credits-lab

Portfolio research project using public and clearly-labeled synthetic data. Modeled estimates, not realized losses. Not a production lending, fraud, or regulatory-compliance system.

---

What it answers

Question	Module
Who should get credit?	Underwriting engine → PD, risk grade, approve / review / decline, credit limit, reason codes
Which payments are risky?	Fraud & payments monitor → fraud score, action (approve / step-up / review / block), expected loss
How much could be lost?	Expected-loss engine → PD × LGD × EAD, by segment, under base and stressed conditions
Can the models be trusted?	Model-risk validation → calibration, drift, segments, Pass / Monitor / Fail verdicts
What if thresholds change?	Policy simulator → 180 precomputed scenarios trading growth against loss and review burden

Architecture

data ingestion → feature engineering → model training + calibration → decisioning
→ expected loss → validation → reporting outputs → React frontend

• Backend: Python 3, pandas, NumPy, scipy, scikit-learn, joblib, pytest. Modular src/ package (data, models, risk, validation, reporting) driven by one orchestrator, src/run_pipeline.py (14 fail-fast phases, deterministic).
• Frontend: React 19 + TypeScript + Tailwind + Radix + Recharts + Vite. Reads only the reconciled static outputs in data/outputs — no hardcoded metrics.
• Code evidence: 30 Python source files, 20 test files, 40 test functions.

Data sources

• Credit: LendingClub accepted loans (public, Kaggle wordsforthewise/lending-club), seeded stratified sample; rejected applicants for approval-funnel context.
• Card payments: Kaggle credit-card fraud (mlg-ulb/creditcardfraud) — real labels + Amount + Time + V1..V28; synthetic, clearly-labeled contextual features attached.
• Macro: FRED public series (unemployment, policy rate, inflation, credit-card delinquency, charge-off) — stress overlays only.
• Stablecoin: fully synthetic, seeded (secondary module).

Leakage is guarded structurally: only application-time credit fields are read; fraud_flag / chargeback_loss are excluded from fraud model features; a data-quality + leakage gate must pass before any output is written.

Key results (from data/outputs)

• 94,999 credit applicants · 69,999 underwriting decisions · 80,000 payment transactions · 2,500 stablecoin transactions · 180 policy scenarios.
• Underwriting champion (logistic scorecard, calibrated): ROC-AUC ≈ 0.695 — verdict Monitor (usable, modest discrimination, disclosed).
• Fraud supervised model: PR-AUC ≈ 0.66, ROC-AUC ≈ 0.97, capture ≈ 0.81 at the review+block boundary — verdict Pass.
• Stablecoin AML-style scorer: discrimination AUC ≈ 0.87 vs. the synthetic label — verdict Pass.

Run it

Requires Python 3 (python3) and Node. Kaggle API token at ~/.kaggle/kaggle.json is needed to regenerate raw data.

# backend: regenerate all 25 reconciled outputs (deterministic)
python3 -m pip install -r requirements.txt
python3 -m src.run_pipeline

# tests
python3 -m pytest -q

# frontend (syncs data/outputs, serves the terminal)
cd frontend && npm install && npm run dev

data/outputs is committed, so the frontend renders immediately without re-running the pipeline.

Assumptions & limitations

• default_flag is a default / severe-delinquency proxy (Charged Off / Default / Late 31–120 / charged-off "does not meet policy"), not pure charge-off default.
• Expected-loss dollars are assumption-driven (LGD/EAD/severity/stress are labeled assumptions), not observed losses from an active book.
• Card contextual features and the entire stablecoin domain are synthetic; the stablecoin label is synthetic.
• amount_zscore_by_account uses each account's full-window mean/std (mild temporal lookahead), disclosed.
• Macro variables inform stress overlays only; no individual-default causality is claimed.
• No real customer data, no real lending decisions, no regulatory-compliance claim.

Deployment

The site lives in frontend/ and deploys on Netlify from this repository (continuous deployment — pushing to main rebuilds and redeploys). Build: base frontend, command npm run build, publish dist.