Credit Risk Modeling Based on XGBoost and Neural Networks: A Complete Practice from Feature Engineering to Strategy Optimization

This article analyzes a complete credit risk modeling project, covering large-scale data preprocessing, XGBoost feature selection, neural network modeling, SHAP interpretability analysis, and approval strategy optimization, providing data-driven risk control decision-making solutions for financial institutions. The project combines the advantages of XGBoost and neural networks to balance risk and return, and achieve an interpretable and implementable risk control system.

The core goal of the project is to develop a machine learning-driven credit risk assessment model to predict customer default probability and support credit decision-making. Based on the American Express Kaggle public dataset (13 months of behavioral data and default labels from April 2017 to April 2018), the business requirement is to maximize expected returns under the premise of controlling default risks, and to formulate differentiated approval strategies to balance conservative loan rejection and aggressive customer acquisition.

Credit risk data contains multi-dimensional fields (behavior, payment, consumption, balance) and has problems of missing values, outliers, and unbalanced distribution. The processing flow includes missing value handling, anomaly detection, and data type conversion; since it contains time-series features, a strategy needs to be designed to convert the 13-month rolling data into static features.

Feature construction strategies:

1. Basic statistical features (mean, standard deviation, etc.) to characterize behavioral stability;
2. Trend features (slope, change rate) to capture behavioral trends;
3. Ratio features (credit utilization rate, repayment rate, etc.) to improve predictive power;
4. Category encoding to handle non-numeric features. A subset of features is selected by calculating feature importance via XGBoost to reduce complexity, reduce overfitting, and improve efficiency.

Two models are trained using an ensemble approach:

• XGBoost: Strong structured data processing capability and interpretable, with stable performance after hyperparameter tuning (learning rate, tree depth, etc.);
• Neural Network: MLP architecture with Dropout regularization and early stopping mechanism to capture complex feature interactions. The fused results form a robust ensemble, balancing interpretability and expressive power.

Financial models need to be interpretable (for regulatory, trust, and debugging needs). SHAP is introduced to quantify the contribution of features to individual predictions, answering:

• Which features have the greatest impact?
• Why a specific customer got a certain score?
• What is the correlation between features and the target variable? This enhances decision transparency and credibility, supporting business communication.

Strategy comparison:

• Conservative strategy: High risk threshold, low default rate but limited returns;
• Aggressive strategy: Low threshold, expanded approval scope but increased losses. Decision-making is assisted by simulating expected returns and risk exposure. Practical suggestions:

1. Attach importance to data quality, invest in data exploration and cleaning in the early stage;
2. Build features with financial meaning in combination with business;
3. Standardize interpretability tools like SHAP;
4. Collaborate with business teams to convert model outputs into executable strategies.