AMEX RiskIQ

14 Problems · 5 Phases · 73 Notebooks

Enterprise Credit Risk Platform, built on real results.

An end-to-end credit risk system on the real Kaggle American Express Default Prediction dataset — probability of default through collections optimization, credit-line management, and executive reporting. Every figure below is either computed live by that problem's own code, or an explicitly labeled ASSUMPTION where the data genuinely has no ground truth. Nothing is rounded up, and the one system that didn't clear its own KPI bar is reported that way, not hidden.

Project Portfolio Prepared By Nandagopal R

View Executive Rollup Dashboard → Browse Source on GitHub Full README
$429,926,252.73
Total real net value / cycle, across 9 production-recommended problems
0.9620
Champion model holdout AUC (Problem 1, XGBoost, measured)
458,913 / 924,621
Training / test customers aggregated (measured)
9 of 14
Problems recommended for production on their own real KPI gates

The 5 phases, in order

Every problem below links straight to its own live, GitHub-Pages-rendered dashboard and its full README. Dollar figures and ratios are copied verbatim from each problem's own measured results — nothing here is a separate estimate.

PHASE 1 Foundation The base PD model and risk-tier map everything else builds on
PROBLEM 01FOUNDATIONAL
Credit Scoring / PD Prediction
Champion model: XGBoost · Holdout AUC 0.9620 · AMEX metric 0.7936
Dashboard · README
PROBLEM 02FOUNDATIONAL
Risk Tier Classification
4 tiers · chi-square p≈0.0 · Cramér's V 0.76
Dashboard · README
PHASE 2 Regulatory & Loss Provisioning IFRS9/CECL reserves, loss severity, earliest default signal
PROBLEM 03RECOMMENDED
Expected Credit Loss (IFRS9 / CECL)
Reserves — IFRS9 $72.4M · CECL $73.1M · flat baseline $67.8M
Dashboard · README
PROBLEM 04FOUNDATIONAL
Delinquency Escalation & Loss Severity
3 LGD tiers — 0.30 / 0.45 / 0.65 · Severe-to-Low default ratio 48.7×
Dashboard · README
PROBLEM 05RECOMMENDED
Early Payment Default Detection
Reliable at 3 statements · retains 96.3% of full-history AUC
Dashboard · README
PHASE 3 Behavioral Intelligence Trailing-window scoring, early warning, roll-rate transitions
PROBLEM 06RECOMMENDED
Dynamic / Behavioral Credit Scoring
Winning window W=3 · retains 99.2% of full-history AUC
Dashboard · README
PROBLEM 07RECOMMENDED
Early Warning System
Alert capture rate 0.4051 · net benefit $2,912,145/cycle
Dashboard · README
PROBLEM 08RECOMMENDED
Roll-Rate Modeling
Severe/Low default ratio 107.2× · both hard-gate KPIs met
Dashboard · README
PHASE 4 Operational Risk Management Collections, credit lines, real-time portfolio monitoring
PROBLEM 09RECOMMENDED
Collections Optimization
Net benefit $341,908,210/cycle — the platform's single largest value-creation figure
Dashboard · README
PROBLEM 10NOT RECOMMENDED
Credit Line Management
KPI target not met (with CI) · net benefit would be $4,108,042.50/cycle if deployed anyway — shown for completeness, not as a recommendation
Dashboard · README
PROBLEM 11RECOMMENDED
Real-Time Portfolio Monitoring
Cohort default-rate lift 2.67× · MCC 0.401
Dashboard · README
PHASE 5 Customer & Business Intelligence Unified profiles, profitability, executive aggregation
PROBLEM 12RECOMMENDED
360° Customer Intelligence
Composite AUC 0.9590 vs. best-single 0.9626 · net benefit $1,275,022/cycle
Dashboard · README
PROBLEM 13RECOMMENDED
Risk-Adjusted Profitability Modeling
Spearman(risk, profitability) -0.9784 · net benefit $58,060,687.58/cycle
Dashboard · README
PROBLEM 14RECOMMENDED
Executive Decision Support Dashboard
BI aggregation across all 13 prior problems · $429,926,252.73/cycle total, 4× independently reproduced
Dashboard · README

Reported honestly, including the miss

Of the 14 problems, 9 are production-recommended value-creation systems summing to $429,926,252.73/cycle, 2 are foundational models with no standalone dollar figure, 1 (Problem 3) is a reserve-accuracy gain kept in its own category, and 1 (Problem 10) is honestly excluded — its real trend-coherence and minimum-tier-population KPI checks did not clear their own bar on the real, measured backtest, even after a documented root-cause investigation and redesign. It ships anyway, in full, for transparency — see its README and CHANGELOG for the full before/after.