Explainable Federated Learning for U.S. State-Level Financial Distress Modeling

Battle Count: 78

The paper focuses on consumer-level financial distress prediction (debt collection agency contact) using survey data, not on market microstructure, asset pricing, or trading strategies. While the FL and XAI techniques could theoretically be adapted for institutional risk modeling, the direct relevance to quantitative trading is minimal. The work is more aligned with consumer credit risk, financial inclusion policy, and regulatory compliance than with trading or portfolio management.

The implementation requires setting up a cross-silo FL infrastructure (Flower-FL framework), designing an 8-layer Highway Network with gating mechanisms, implementing class weighting for imbalanced data, and integrating two XAI methods (SHAP and Owen values). The partial participation strategy and FedAvg aggregation add distributed systems complexity. However, the use of established frameworks (Flower-FL, PyTorch) and publicly available data reduces some barriers. The main complexity lies in coordinating 51 clients, managing communication rounds, and interpreting the dual XAI outputs for categorical survey features.

Open-source code is available on GitHub (https://github.com/brains-group/xfl-for-loan-eligibility). The NFCS dataset is publicly available from the FINRA Foundation. Detailed hyperparameters (class weight factor 1.1835, 200 training rounds, 12/51 partial participation, Adam optimizer, early stopping) are provided. The Flower-FL framework is used. Minor gap: exact random seeds and full preprocessing pipeline details could be more explicit.