Decoupling Household Financial Vulnerability in China : A Heterogeneous Stacking Ensemble Approach with Interpretability-Driven Risk Attribution

Household debt risk has evolved into a critical non-linear variable affecting systemic financial stability. Traditional linear models and single machine learning algorithms often struggle to capture the complex, high-dimensional interaction effects inherent in micro-financial data. To address this, this study proposes a heterogeneous stacking ensemble framework that explicitly integrates the diverse inductive biases of gradient boosting decision trees (LightGBM), random forests, and deep neural networks, utilizing logistic regression as the meta-learner. A structured design component of this framework is a semantic-aware feature reshaping mechanism, which organizes tabular financial data into economically meaningful clusters, enabling convolutional neural networks (CNN) to extract local structural interaction patterns. Validated on the China Family Panel Studies (CFPS) dataset (N = 35, 636
), the system achieves an AUC of 0.912 and a Recall of 0.854, demonstrating improved predictive performance over state-of-the-art single benchmarks. Furthermore, utilizing SHAP game-theoretic analysis, we identify a salient predictive pattern, revealing that a structural mismatch—asset-rich but cash-poor—is a prominent risk indicator of household default. Within the Chinese household sector examined here, this study contributes a transparent decision support tool whose findings suggest that liquidity-based metrics may complement traditional collateral-based risk assessment; calibration and temporal out-of-sample analyses further support the framework’s robustness.