Credit Risk Monitor

A credit risk monitoring framework for consumer installment lenders operating at the Regional Finance tier (mid-size, near-prime/subprime consumer lenders). It combines public mortgage-lending data (HMDA) and consumer complaint data (CFPB), enriched with FRED macroeconomic indicators, into a risk segmentation and champion/challenger modeling pipeline, a fair-lending parity screen, a population-stability drift monitor, and an interactive Texas county heatmap all surfaced in a Streamlit dashboard.

This is a portfolio/analytical project, not a production underwriting system. Nothing here should be used, or presented, as an actual credit decisioning tool. See CLAUDE.md for full data provenance and project rules.

Dataset Sources

Dataset	Description	Source
HMDA 2022 LAR (Texas)	Home Mortgage Disclosure Act Loan Application Register, filtered to Texas, originated/denied applications	ffiec.cfpb.gov/data-publication
CFPB Consumer Complaints	Payday/personal/consumer installment loan complaints	consumerfinance.gov/data-research/consumer-complaints
FRED (DRCCLACBS, CORCACBS)	Delinquency rate on credit card loans / charge-off rate on consumer loans, all commercial banks	fred.stlouisfed.org
TX county boundaries	Census-derived, FIPS-keyed county GeoJSON, used for the choropleth map	plotly/datasets

HMDA covers mortgage lending, not installment/payday loans — it's used here as a regional proxy for underwriting and fair-lending patterns in Texas, not as direct data on any specific lender's own loan book.

Folder Structure

credit-risk-monitor/
├── data/
│   ├── raw/            # Original source CSVs + cached TX county GeoJSON
│   └── processed/      # Cleaned/engineered datasets (gitignored, regenerable)
├── src/
│   ├── data_loader.py  # Load + clean HMDA and CFPB raw data
│   ├── fred_loader.py  # Pull FRED macro series
│   ├── features.py     # Risk segmentation and fair-lending feature engineering
│   ├── models.py       # Champion (Logistic Regression) / challenger (XGBoost) pipeline
│   ├── monitor.py      # PSI drift monitoring + cohort delinquency projection
│   ├── fair_lending.py # Four-fifths rule disparity screening
│   └── choropleth.py   # Interactive Texas county approval-rate heatmap
├── models/              # Trained model artifacts (gitignored)
├── dashboard/
│   └── app.py           # Streamlit dashboard
├── notebooks/            # Exploratory analysis
├── tests/
├── outputs/              # Generated charts, reports, evaluation CSVs (gitignored)
├── .env.example          # FRED_API_KEY template
├── requirements.txt
└── CLAUDE.md              # Full project context and rules

Setup

git clone https://github.com/Rohanrsp14/Credit-Risk-Monitor.git
cd Credit-Risk-Monitor

python -m venv venv
venv\Scripts\activate        # Windows
# source venv/bin/activate   # macOS/Linux

pip install -r requirements.txt

cp .env.example .env         # then fill in your FRED_API_KEY

Place the raw source CSVs in data/raw/ (hmda_2022_tx.csv, cfpb_complaints.csv — not included in the repo; download from the sources above), then run the pipeline in order:

python src/data_loader.py
python src/fred_loader.py
python src/features.py
python src/models.py
python src/monitor.py
python src/fair_lending.py
python src/choropleth.py

Then launch the dashboard:

streamlit run dashboard/app.py

Dashboard

Model Performance

Cohort Tracking

Fair Lending

Texas Heatmap

Regulatory Framework Context

This project borrows concepts from real bank model-risk and fair-lending practice to give the framework realistic structure — it does not implement or certify compliance with any of the following:

• SR 11-7 (Federal Reserve / OCC Supervisory Guidance on Model Risk Management) — motivates the champion/challenger structure, independent evaluation metrics (AUC, Gini, KS), and PSI-based ongoing monitoring in src/models.py and src/monitor.py. A real SR 11-7 program requires independent validation, documented conceptual soundness review, and effective challenge — none of which this project provides.
• ECOA / Regulation B (Equal Credit Opportunity Act) — motivates the four-fifths disparity screen in src/fair_lending.py. Note that src/models.py includes an is_minority feature in the model per this project's own build spec; using protected-class status as a decisioning input would violate ECOA/Reg B in a real underwriting system. This is only present here as an analytical/portfolio exercise.
• CFPB (Consumer Financial Protection Bureau) — both the HMDA and complaints datasets are sourced from the CFPB, which also supervises fair lending and UDAAP compliance for consumer lenders at this tier.

Never-Do / Always-Do Rules

See CLAUDE.md for the full list — most importantly: no synthetic data, no hardcoded secrets, no overclaiming model accuracy, and every output must cite its data source.