Bioinspired Deep Neural Networks for Predicting Income-Reporting Discontinuities in the Chilean Student Loan Program.

Area of Science:

• Credit Risk Management
• Machine Learning Applications
• Financial Data Analysis

Background:

• Student loan programs face challenges with income reporting discontinuities.
• Accurate prediction of non-compliance is crucial for effective credit risk management.
• Existing machine learning models lack methodologically robust evaluations integrating imputation and validation.

Purpose of the Study:

• To develop and evaluate a robust pipeline for discontinuity prediction in student loan income reporting.
• To compare a bioinspired Deep Neural Network (DNN) against a Random Forest (RF) classifier.
• To provide a replicable methodology for risk analysis in student credit.

Main Methods:

• Implemented a pipeline with MissForest imputation and SMOTE-based balancing on 22,303 records.
• Conducted a comparative assessment of a bioinspired DNN and a Random Forest classifier.
• Utilized 35 stratified partitions for repeated cross-validation.

Main Results:

• The bioinspired DNN significantly outperformed the RF classifier across key metrics (AUC, F1-score).
• DNN demonstrated superior performance with AUC of 0.9991 vs 0.9709 and F1-score of 0.9966 vs 0.9497.
• Interpretability analysis revealed financial variables as key predictors, with minimal influence from demographic data.

Conclusions:

• The bioinspired DNN offers a highly accurate and stable method for predicting income reporting discontinuities.
• The developed methodology enhances credit risk management practices in student lending.
• Financial data holds greater predictive power than demographic data for this specific task.