LoanSight - Credit Risk Engine

Inspiration

Banks lose billions every year approving loans for applicants who default. Traditional credit checks are slow and opaque. We wanted to build a fast, explainable system that gives loan officers a clear decision with reasoning.

What it does

LoanSight scores any loan applicant in real-time and returns:

APPROVE / REJECT / REVIEW decision
Default probability percentage
FICO-style risk score (300–850)
Top risk factors explaining the decision
What-if analysis: change one feature and see how risk shifts instantly

How we built it

Trained an XGBoost pipeline on 150,000 real credit records (Give Me Some Credit dataset)
Built a data validation firewall to catch bad/fake inputs before prediction
Added missingness indicators as features (missing income is itself a risk signal)
Wrapped everything in a clean Streamlit UI with explainability charts

Challenges we ran into

Missing data handling: ~20% of income values were missing solved with MNAR indicators + median imputation inside the pipeline
Threshold optimization: default rate is only 7%, so we used a cost matrix (FN costs 5x more than FP) to set the decision threshold at 0.25
Getting the UI to pass the right feature columns to the trained pipeline

Accomplishments that we're proud of

End-to-end modular system: validation → prediction → explanation in one click
What-if analysis that updates live as you sweep any feature
Clean separation of concerns: preprocess / predict / explain are fully independent