Real-Time Loan Default Risk Prediction System: Multi-Model Comparison and Application of SHAP Interpretability in Financial Risk Control

This article introduces a machine learning system compliant with banking regulatory standards for real-time loan default risk prediction. The project compares three models: logistic regression, XGBoost, and neural networks, integrates SHAP interpretability analysis, balances prediction performance and regulatory transparency, and helps credit teams understand the basis for risk decisions. The project is open-source and has been deployed as an interactive application with clear commercial value.

Traditional credit scoring models struggle to capture complex non-linear features of default risk due to linear assumptions. Financial institutions face the contradiction between 'maintaining regulatory transparency' and 'improving prediction accuracy'. The core problem of the project is to answer 'whether the applicant will repay on time', which needs to meet: accurately identifying high-risk applicants, providing clear decision-making basis, being easy to use for non-technical users, and responding in real-time to the approval process.

The project trained and compared three models, with the following results:

Model	Accuracy	ROC-AUC	Default Recall Rate	Evaluation Conclusion
Logistic Regression	86.84%	0.745	0.50	Baseline interpretable model
XGBoost	100%	1.0	1.0	Optimal performance
MLP Neural Network	99.94%	0.9995	1.0	Near optimal
The author confirmed no data leakage in XGBoost/MLP's high scores, and the performance of logistic regression is consistent with the limitations of linear models in handling non-linear data. Ensemble trees and neural networks can capture interactive features and non-linear relationships missed by traditional models.

Through SHAP analysis and XGBoost feature importance, core default signals were identified:

• Interest Rate Related: Interest rate spread (primary factor), loan interest rate, upfront fees
• Borrower Features: 45-54 age group, income level, debt-to-income ratio
• Loan Structure: Loan limit, property value, loan-to-value ratio, credit score

Interpretability solutions:

1. SHAP LinearExplainer (for logistic regression): Provides marginal contributions of features, aligning with the intuition of linear models.
2. XGBoost native feature importance: Based on gain and split times, shows internal decision logic. The results of both are consistent, enhancing credibility.

The project adopted a systematic data processing approach:

• Missing Values: Categorical variables filled with mode, numerical variables filled with mean (completed before split to prevent leakage)
• Outliers: IQR tail capping (needs calibration to avoid feature skewness)
• Class Imbalance: SMOTE oversampling for minority class
• Feature Scaling: StandardScaler used for logistic regression/MLP, none for XGBoost
• Encoding: One-hot encoding (gender, loan type, etc.), binary encoding (8 features like loan limit), all executed after split.

The XGBoost model has been deployed as a Render application (online demo: https://loan-default-risk-3yo4.onrender.com/docs; local run: streamlit run app.py). The UI design supports credit officers to input information and get real-time default probability, risk score, and feature contributions.

Commercial value calculation (assuming 10k monthly applications, average loan £15k, default rate 24%):

Scenario	Annual Default Loss
No model (full approval)	£43,200,000
Manual review (70% interception rate)	£12,960,000
Logistic regression (50% recall)	~£21,600,000
XGBoost (100% recall)	Theoretically close to zero
V2 goal: Increase default recall rate to 75% to maximize commercial value.

Technical Highlights:

1. Model selection trade-off: Systematically compares models of different complexities to understand the source of performance improvement.
2. Interpretability first: SHAP was incorporated from the initial stage, not as an afterthought.
3. Data quality engineering: Systematic handling of missing values, outliers, etc.
4. Anti-data leakage: Preprocessing steps (encoding, scaling) executed after split.
5. End-to-end deployment: Complete process from data exploration to production deployment.

Summary: The project demonstrates the responsible application of AI in financial risk control, balancing performance and transparency, and provides a reference case for practitioners. Open-source implementation and online demo lower the learning threshold.