01
OPTI-RECOURSE
Credit-risk prediction platform that explains every rejection and tells the borrower how to flip it.
- Role
- Solo build: modelling, explainability, app
- Stack
- Python · Pandas · NumPy
- Year
- 2025
- Status
- Published · Live
Context
Credit rejections are usually a black box. A borrower gets a no, with no indication of what would have made it a yes, and the lender carries a model nobody in the business can interrogate.
That leaves two problems in one system: applicants have no route back, and risk teams cannot defend a decision when a regulator, an analyst or a customer asks why. OPTI-Recourse was built to close both gaps with the same set of explanations.
Approach
- 01
Built and compared three classifiers (XGBoost, Logistic Regression and Random Forest) on a held-out test set with cross-validation.
- 02
Tuned hyperparameters with Optuna, targeting recall so genuinely risky applicants aren't missed.
- 03
Layered SHAP for global, portfolio-level feature attribution and LIME for individual predictions.
- 04
Turned the explanations into recourse: the specific feature changes that would flip a rejection.
- 05
Shipped it as an interactive Streamlit dashboard for real-time risk scoring.
- 06
Wrote up the methodology; peer-reviewed and published in IJSREM.
Results
- more risky borrowers caught
- +12%
- models compared head to head
- 3
- ways it explains a rejection
- 2
- peer-reviewed publication
- 1
Stack
- Python
- Pandas
- NumPy
- Scikit-learn
- XGBoost
- Optuna
- SHAP
- LIME
- Streamlit
DATA DESK
AD CAMPAIGN DATA INTEGRITY