FedIncome: Federated Learning for Income Estimation in Digital Lending Under Data Sovereignty Constraints (arXiv (Cornell University))

Summary

• The FedIncome framework uses over one million LendingClub loans partitioned into 50 state-level clients to simulate a heterogeneous lending consortium.p.1
• The best federated model achieves an out-of-time R2 of 0.608, compared with 0.619 for a pooled centralised benchmark.p.1
• Data-scarce institutions with fewer than 1,807 training observations obtain average out-of-time R2 improvements of 3.8 percentage points relative to the pooled centralised benchmark.p.3
• Replacing reported income with FedNova predictions increases simulated approval rates by 4.6 percentage points in the source-verified group and 7.8 percentage points in the not-verified group.p.22
• Relative to local-only training, federation improves out-of-time R2 across all client sample-size groups, with gains of 0.152 for low-sample clients and 0.027 for high-sample clients.p.33

Key insights

1. The empirical crossover between federated and pooled centralised model performance occurs at approximately 4,790 training observations in the LendingClub setting.p.1
2. The Residual MLP base architecture achieves an out-of-time R2 of 0.641 under centralised model selection.p.16
3. FedNova attains an out-of-time RMSE of $32,869 compared with $32,394 for the centralised benchmark.p.17
4. Estimated DTI thresholds range from 11.2% to 46.7% across state-income cells, with mean thresholds rising from 22.7% for low-income to 28.9% for high-income borrowers.p.21
5. Replacing reported income with FedNova predictions increases observed default rates among approved loans by 0.29 percentage points for source-verified and 0.31 percentage points for not-verified borrowers.p.22
6. FedNova achieves higher out-of-time R2 than local-only models in 94.1% of low-sample states and 100.0% of mid-sample states.p.33