Leveraging Multiview Data Through Discrete and Regularized Deep Learning for Dynamic Financial Risk Prediction
提出一种离散正则化深度学习（DRDL）方法，利用多视图数据提升动态金融风险预测性能，在模型和应用层面均优于现有方法，并具备可解释、冗余过滤、功能解缠和单调性等实用优势。
Given the dramatic surge of demand for predictive insights into the dynamics of financial risk and the rich, yet entangled, information brought by proliferating multiview data, we propose a discrete and regularized deep learning (DRDL) method to better leverage such multiview data for dynamic financial risk prediction. Empirical evaluation demonstrates advantages of DRDL over benchmarked classic and state-of-the-art methods at both the model level (time-to-risk and out-of-time prediction performance) and the application level (identification and profitability performance). Besides performance gains, DRDL offers distinctive practical advantages. First, it enables explicit and controllable factor-level representations, allowing practitioners to inspect and regulate how cross-view signals are encoded. Second, it offers unique advantages in explicitly and precisely filtering out redundant information while extracting complementary information across heterogeneous data sources, allowing practitioners to better understand which unique informational components drive risk predictions. Third, it offers a practically viable and empirically effective way to promote functional disentanglement within a discrete and structured latent space. Fourth, it supports both time-wise and instance-wise monotonicity, aligning predictions with the cumulative and irreversible nature of financial risk escalation, which may be particularly valuable in risk monitoring and governance contexts.