Risk-Sensitive Machine Learning for Financial Decision Modeling Under Imbalanced Data: Evidence from Bank...

Area of Science:

• Machine Learning
• Data Science
• Financial Analytics

Background:

• Bank telemarketing campaigns face low subscription rates due to customer differences and imbalanced datasets.
• Predictive modeling for telemarketing outcomes is challenging because of severe class imbalance.

Purpose of the Study:

• To enhance the prediction of bank telemarketing outcomes using a data-driven approach.
• To integrate synthetic minority oversampling and cost-sensitive learning for improved predictive accuracy.

Main Methods:

• Utilized the Portuguese Bank Marketing dataset (41,188 instances, 11.3% positive response).
• Evaluated eight machine learning models (Logistic Regression, Decision Tree, Random Forest, Ensemble methods) using cross-validation.
• Applied synthetic minority oversampling and cost-sensitive learning techniques.

Main Results:

• Ensemble models (CatBoost, XGBoost, LightGBM) outperformed traditional baselines.
• Achieved significant gains in minority-class recall and overall discrimination.
• The best model reached an F1-score of 0.540, positive class recall of 0.812, and ROC-AUC of 0.908.

Conclusions:

• Combining resampling strategies with cost-sensitive optimization offers a robust method for imbalanced telemarketing data.
• SHAP analysis identified key predictors: campaign duration, previous contact outcomes, and macroeconomic indicators.
• This approach supports reproducible, data-driven financial decision-making by addressing minority-class identification challenges.