An Advanced AI-Driven Risk Assessment Framework for U.S. Banking Institutions: Integrating Predictive Financial Analytics, Regulatory-Aware Governance and Cybersecurity Risk Intelligence

Abstract

This study proposes an advanced AI-driven risk assessment framework for U.S. banking institutions by integrating predictive financial analytics, regulatory-aware governance, and cybersecurity risk intelligence within a unified decision-support model. The research responds to the increasing complexity of the U.S. banking environment, where financial instability, regulatory pressure, digital fraud, third-party dependency, and cyber threats require more adaptive and intelligent risk-management systems. Existing banking risk models often operate in isolated domains, limiting their ability to provide real-time, cross-functional, and explainable risk insights. A secondary and simulation-based dataset was developed using publicly available U.S. banking indicators, Federal Reserve financial stability variables, FDIC bank performance data, anonymized transactional risk records, cyber threat intelligence feeds, fraud-alert logs, and regulatory compliance indicators. The final dataset consisted of 125,000 banking-risk observations categorized into five major risk classes: credit default risk, liquidity stress, operational failure, regulatory non-compliance, and cybersecurity intrusion. The proposed framework was implemented using Python 3.11, Jupyter Notebook, Scikit-learn, TensorFlow/Keras, XGBoost, SHAP, Pandas, NumPy, Matplotlib, SQL-based storage, and Power BI visualization. The methodological process included data cleaning, missing-value treatment, z-score normalization, SMOTE-based class balancing, feature engineering, correlation filtering, and an 80:20 train–test split with five-fold cross-validation. Several machine-learning and deep-learning models were evaluated, including Logistic Regression, Random Forest, XGBoost, Long Short-Term Memory networks, and a proposed hybrid XGBoost–LSTM ensemble. SHAP explainability was incorporated to improve model transparency, while the regulatory-governance layer mapped AI-generated risk scores against U.S. supervisory expectations related to model validation, cybersecurity controls, operational resilience, auditability, and compliance traceability. The experimental results demonstrate that the proposed hybrid XGBoost–LSTM ensemble outperformed all baseline models, achieving 96.4% accuracy, 95.8% precision, 96.1% recall, 95.9% F1-score, and 0.982 ROC-AUC. Compared with Logistic Regression, the proposed framework improved classification accuracy by 14.7%, reduced false-risk alerts by 31.6%, and decreased average risk-detection latency from 2.8 seconds to 0.9 seconds. It also enhanced regulatory-risk scoring consistency by 22.4% and cybersecurity incident prioritization by 27.8%. These findings confirm that integrated AI-driven risk assessment can strengthen early-warning capability, compliance readiness, cyber-resilience, and strategic decision-making across U.S. banking institutions.