JenniferAnnKok/ACRIE

🏦 Home Credit Default Risk Predictor

An interactive credit scoring app built with XGBoost, SMOTE, and SHAP explainability — part of the Investec Portfolio Project.

How it works

1. Fill in applicant details across the four input tabs:

   • Application — income, loan amount, personal details
   • Credit scores — external bureau scores (EXT_SOURCE_1/2/3)
   • Bureau & history — bureau loan aggregates and previous applications
   • Instalments — payment behaviour history

2. Click Predict default risk

3. The app returns:

   • A probability gauge showing the estimated default probability
   • A SHAP waterfall chart explaining which features drove the score
   • A decision summary table (APPROVE / REJECT at the cost-optimal threshold)

Model details

Property	Value
Algorithm	XGBoost (hist tree method)
Imbalance handling	SMOTE (sampling_strategy=0.3)
Hyperparameter search	RandomizedSearchCV (15 iter, 3-fold CV)
Evaluation metric	ROC-AUC
Target threshold	0.27 (cost-optimal, not default 0.50)
Explainability	TreeSHAP via shap library

Using your own trained model

Upload these three files (produced by the training pipeline) to the Space's file system:

home_credit_xgb_model.pkl      ← joblib.dump(best_xgb, ...)
home_credit_threshold.pkl      ← joblib.dump(cost_optimal_thresh, ...)
home_credit_features.pkl       ← joblib.dump(X_train.columns.tolist(), ...)

The app detects them automatically on startup. If they are absent it falls back to a demo model trained on synthetic data so the Space stays functional.

Threshold logic

The cost-optimal threshold (default 0.27) minimises expected credit loss under:

• LGD = 45% of loan value (loss given default)
• Opportunity cost = 8% annual interest (rejected good client)
• False negatives are ~5× more costly than false positives at median loan size

Data sources (training)

Table	Rows	Purpose
application_train.csv	307 511	Main application features
bureau.csv	1 716 428	External credit history
previous_application.csv	1 670 214	Past Home Credit applications
installments_payments.csv	13 605 401	Repayment behaviour (3M sampled)