Enhancing Transparency in Asset-Backed Securities: A Deep Learning Approach for Automated Risk Assessment and Regulatory Compliance

Abstract

The 2008 financial crisis exposed critical transparency deficiencies in asset-backed securities markets, prompting regulatory reforms mandating asset-level disclosure. This research develops an automated risk assessment framework combining deep neural networks with SHAP explainability techniques to address the regulatory technology gap in processing large-scale securitization data. The framework processes Schedule AL disclosures from SEC Electronic Data Gathering, Analysis, and Retrieval (EDGAR) filings, extracting loan-level and pool-level features to predict default risk while providing interpretable explanations for each assessment. Empirical validation on 450,382 mortgages from 50 residential mortgage-backed securities transactions demonstrates superior performance with an AUC-ROC of 0.883, outperforming XGBoost by 2.7 percentage points while maintaining complete transparency through feature attribution. Case studies illustrate the practical applications of detecting underwriting quality deterioration and geographic risk concentration, thereby supporting regulatory compliance monitoring and investor protection objectives.