FSL-BDP: Federated Survival Learning with Bayesian Differential Privacy for Credit Risk Modeling

Battle Count: 73

The paper is primarily focused on credit risk modeling and lending decision support rather than quantitative trading. However, there are tangential connections: (1) survival/time-to-default modeling could inform credit derivative pricing and structured product risk assessment; (2) the federated learning + privacy framework could be adapted for multi-institutional risk aggregation relevant to systemic risk monitoring; (3) the finding about privacy mechanism ranking reversal under federation is methodologically relevant for any multi-party financial ML system. The direct applicability to trading strategies, market making, or algorithmic execution is minimal.

High complexity due to the integration of multiple advanced components: (1) discrete-time survival analysis with neural network hazard parameterization, (2) federated learning protocol with FedAvg aggregation across 27-32 clients, (3) Bayesian differential privacy with leave-one-out gradient estimation, Monte-Carlo sensitivity estimation, and Rényi divergence-based privacy accounting, (4) handling of non-IID data distributions across clients, (5) proper censoring treatment in survival likelihood, (6) privacy budget composition across rounds and epochs. The Bayesian DP algorithm (Algorithm 2) requires computing per-sample gradients, constructing neighboring batches, and evaluating Rényi divergence at multiple orders λ. Requires careful hyperparameter tuning for privacy-utility trade-off.

The paper specifies fixed random seeds (seed=42), deterministic CUDA settings, 5 independent runs with standard deviations reported, detailed hyperparameters (Adam optimizer η=0.001, batch size B=32, R=10 rounds, E=5 local epochs, C=1.0, σ calibrated for ε∈{0.5,1.0,2.0,10.0}, δ=10^-5, M=10 Monte-Carlo samples). Implementation in PyTorch 2.0 on NVIDIA T4 GPUs. However, no GitHub repository is mentioned, and the Bayesian DP algorithm details are provided in pseudocode (Algorithm 1 and 2). Datasets (LendingClub, SBA, Bondora) are publicly available.