2025 • AI & ML / Research

Bank Failure Prediction

Explore Research CCSC Poster

0 %

XGBoost AUC

Classification performance

Quarter Lead Time

Early warning prediction window

Key Indicators

Selected features after RFE

CCSC

Conference

Presented April 2025

Research Overview

This collaborative research project focuses on improving early warning systems for bank failures using machine learning techniques. Presented at the CCSC Central Plains Conference in April 2025, our work investigates how advanced ML models can identify early warning signs of financial instability.

Traditional statistical models face significant challenges due to sparse failure events and highly non-linear financial indicators. Our approach examines how machine learning can provide regulators with valuable lead time for intervention before crisis points emerge.

The research team includes Khalid Mohammed, Coleman Pagac, Rediet Ayalew, and Braedon Stapelman, working under the supervision of Dr. Eric Manley and Dr. Sean Severe at Drake University.

HEALTH Score Framework

Capital Adequacy

Measures financial resilience through the ratio of equity minus goodwill to total assets.

Asset Quality

Evaluates lending portfolio stability and exposure to non-performing assets.

Liquidity

Assesses ability to meet short-term obligations with SHAP importance of 0.23.

Profitability

Examines sustainable earnings capacity and capital adequacy metrics (SHAP 0.19).

Methodology

Traditional Methods

XGBoost AUC 0.93

Random Forest Ensemble baseline

Logistic Regression Interpretability

SVM Margin classifier

Deep Learning

Transformer AUC 0.892

LSTM Sequential patterns

GRU Temporal modeling

CNN-RNN Hybrid Multi-scale features

Feature Engineering

SMOTE Class balancing

RFE Feature selection

SHAP Explainability

Z-Score Normalization

Research Challenges

Severe Class Imbalance

Challenge: Bank failures are rare events, creating significant training challenges with heavily skewed class distributions in historical FDIC data.

Solution: Addressed with weighted loss functions and SMOTE augmentation techniques. Emphasized lower HEALTH ranges via exponential rescaling to improve model sensitivity to early signs of deterioration.

Temporal Dependencies

Challenge: Financial indicators show complex time-varying relationships that simple models cannot capture, requiring sequential modeling approaches.

Solution: Used a fixed four-quarter input window (quarters 6-9 prior to prediction) allowing predictions six quarters in advance. Transformer architectures excel at capturing these deterioration patterns.

Regulatory Interpretability

Challenge: Balancing predictive power with regulatory transparency requirements, as black-box models face adoption barriers in financial oversight.

Solution: Recursive feature elimination combined with SHAP value analysis helps select high-impact, interpretable features. The HEALTH framework supports both risk flagging and continuous monitoring with clear explanations.

Key Findings

"Our findings validate ML architectures as valuable tools for bank failure prediction. ML-enhanced systems can identify at-risk institutions up to six quarters before traditional warning signs appear."

XGBoost delivers superior classification (AUC 0.93) for early warning detection

Transformer architectures (AUC 0.892) excel at capturing deterioration patterns

Liquidity ratios and capital adequacy emerge as critical predictors

The HEALTH framework enables both risk flagging and continuous monitoring

Explore More

CCSC Poster

Conference Presentation

→

FDIC Data

Financial datasets

→

Drake CS

Research department

→

All Projects

Previous Project Algebraic Compiler

Next Project ISEAGE CTF