FedIncome: Federated Learning for Income Estimation in Digital Lending Under Data Sovereignty Constraints

Sultan Amed; Tanmay Sen; Sayantan Banerjee

stat.MLSep 23, 2026

FedIncome: Federated Learning for Income Estimation in Digital Lending Under Data Sovereignty Constraints

Authors: Sultan Amed, Tanmay Sen, Sayantan Banerjee

Organizations: OM & QT Area, Indian Institute of Management Indore, M.P. 453556, India. · SQC & OR Unit, Indian Statistical Institute Kolkata, W.B. 700108, India

Abstract

Verified income is often unavailable in digital loan applications, forcing lenders to rely on reported income and potentially leading to over-lending, overly conservative offers, or rejection of creditworthy applicants. Cross-institutional data-sharing constraints make this problem especially difficult for smaller lenders with limited training data. We introduce FedIncome, a federated learning framework for income estimation that enables institutions to train a shared model without pooling raw borrower records. Using more than one million LendingClub loans partitioned into $50$ state-level clients, we simulate a heterogeneous lending consortium. The best federated model achieves out-of-time $R^2=0.608$ , compared with $0.619$ for a pooled centralised benchmark. Small-sample clients obtain an average out-of-time $R^2$ improvement of $3.8$ percentage points relative to the pooled centralised benchmark, while the fitted client-level relationship places the empirical crossover at approximately $4,790$ training observations in this setting. When pooling is infeasible and the relevant alternative is local-only training, federation improves out-of-time performance across all sample-size groups, with the largest gains for data-scarce clients. We also combine federated income estimates with state- and income-specific debt-to-income thresholds. In a retrospective decision analysis, replacing reported income with the federated estimate increases simulated approval rates with only modest changes in observed default rates. FedIncome supports collaborative learning under data-locality constraints with little aggregate loss relative to pooled training and larger gains relative to local-only estimation.

Figures & tables

Figure 1: AI-led credit decision support system for digital lending.

Figure 2: FedIncome framework : federated income estimation across $K=50$ clients under data-local training. Borrower-level records remain at the participating clients, while model updates are exchanged with the aggregation server during training.

Figure 3: Loan amount calibration framework . Federated income estimates

\hat{Y}_{\text{FedIncome}}

are translated into DTI-constrained loan offers. Income determines the maximum serviceable installment, product parameters (

R,t

) define repayment structure, and the annuity formula yields

L_{\max}

subject to the selected DTI constraint.

Category	N	Representative Features
Credit History & Delinquency	14	Delinquencies in past 2 years, months since last delinquency, accounts 90+ days past due, percentage of accounts never delinquent
Account Counts	12	Number of open accounts, total accounts, accounts opened in past 24 months, trade lines opened in past 12 months
Revolving Credit	15	Revolving balance, bank-card utilization ratio, available credit on bank cards, total revolving credit limit, active revolving accounts
Installment Credit	8	Total installment balance, installment utilization, number of installment accounts, installment accounts opened in past 24 months, age of oldest installment account
Credit Inquiries	5	Inquiries in past 6 months, inquiries in past 12 months, finance inquiries, months since most recent inquiry
Balances & Limits	8	Total current balance, average current balance, total high credit limit, total balance excluding mortgage, maximum balance on bank cards

Table 1: Feature categories used in the income-estimation model.

Panel A: Verification Tiers	N Loans	% Total	Mean Income ($)	Default Rate
Verified (Learning)	493,820	29.70%	78,432	20.85%
Source Verified	664,176	39.97%	81,872	17.87%
Not Verified	503,293	30.33%	71,967	12.58%
Total	1,661,289	100%

Panel B: Temporal Split †	N Loans	Period

Table 2: Sample description.

Model	Dataset	$R^{2}$	RMSE ($)	MAPE (%)	MBE ($)
Ridge Regression	Train	0.497	33,313	34.99	0
	Test	0.504	32,750	34.93	+183
	OOT	0.563	34,697	36.79	$-$ 1,320
Lasso Regression	Train	0.497	33,313	34.99	0
	Test	0.504	32,750	34.93	+183
	OOT	0.563	34,697	36.80	$-$ 1,318

Table 3: Centralized benchmark results.

		In-Time Test			Out-of-Time
Model	Type	$R^{2}$	RMSE ($)	MAPE (%)	$R^{2}$	RMSE ($)	MAPE (%)
Centralized	Benchmark	0.551	31,554	32.83	0.619	32,394	35.59
FedNova	Federated	0.543	31,834	30.29	0.608	32,869	31.83
FedAvg	Federated	0.537	32,063	30.28	0.604	33,034	31.81
FedProx	Federated	0.525	32,460	30.43	0.591	33,564	32.03
FedAdam	Federated	0.534	32,164	31.06	0.578	34,064	37.96

Table 4: Performance comparison: centralized versus federated models.

	Low Income		Mid Income		High Income
Model	MAPE	MBE	MAPE	MBE	MAPE	MBE
Centralized	53.7%	$13,917	24.9%	$5,542	25.8%	$-$ $16,140
FedNova	46.2%	$11,206	20.2%	$684	27.0%	$-$ $26,424
$\Delta$ MAPE	$-$ 7.5 pp		$-$ 4.7 pp		$+$ 1.2 pp
Income terciles defined within the OOT sample. pp = percentage points.

Table 5: Decomposition of OOT prediction error by income tercile.

Figure 4: Convergence of the five federated aggregation procedures relative to centralized training. Configuration:

K=50

clients,

T=10

communication rounds,

E=10

local epochs, and 30% client participation per round.

Figure 5: Client-level performance of FedNova and the pooled centralized benchmark across the 50 state-level clients.

			In-Time Test			Out-of-Time
Category	States	Avg. $n$	$\Delta R^{2}$	Win%	$\Delta$ MAPE	$\Delta R^{2}$	Win%	$\Delta$ MAPE
Low ( $n\leq 1{,}807$ )	17	936	+0.021	70.6%	+3.73	+0.038	76.5%	+4.55
Mid ( $1{,}807<n\leq 5{,}948$ )	16	3,800	+0.005	56.2%	+2.92	$-$ 0.001	56.2%	+4.35
High ( $n>5{,}948$ )	17	13,467	$-$ 0.003	52.9%	+2.62	$-$ 0.009	35.3%	+3.72
$\Delta R^{2}$ = FedNova $-$ Centralized; positive values favor FedNova.
$\Delta$ MAPE = Centralized $-$ FedNova in percentage points; positive values favor FedNova.

Table 6: Performance by client training sample size.

Figure 6: Client-level federation–centralization crossover. Panel (a) plots OOT prediction error

(1-R^{2})

against client training sample size for the pooled centralized benchmark and FedNova. Panel (b) plots the corresponding FedNova

R^{2}

gain against local sample size. The fitted relationship crosses zero at approximately 4,790 training observations in the present experiment. Each point represents one of the 50 state-level clients.

Stratum	Mean	Median	Std Dev	Min	Max
Low Income	22.7	22.1	5.8	11.2	38.4
Mid Income	25.3	24.8	6.2	13.7	42.1
High Income	28.9	28.1	7.5	16.3	46.7
All Strata	25.6	25.0	7.1	11.2	46.7

Table 7: Summary of estimated DTI thresholds (%).

Figure 7: DTI distributions computed from observed or reported income and from FedNova-predicted income across verification tiers. For the verified tier, the observed verified income is the reference. For source-verified and not-verified borrowers, the comparison is against reported income and should not be interpreted as a comparison with latent true income.

Figure 8: Observed default rate by DTI bucket under observed or reported income and under FedNova-predicted income. The comparison describes how the empirical relationship between DTI and subsequent default changes when the income input is replaced by the model prediction.

Metric	Source Verified	Not Verified
Panel A: Sample
Total Loans	664,176	503,293
Panel B: Approval Decisions
Approved (Reported Income)	528,556 (79.6%)	391,439 (77.8%)
Approved (FedNova)	552,826 (83.2%)	421,779 (83.8%)
Approval Rate Change	+4.6 pp	+7.8 pp

Table 8: Approval and observed-default outcomes under reported and FedNova-predicted income.

Explore similar work

Differentially private federated learning with Byzantine-robust aggregation: A cross-domain framework for secure model training in banking and healthcare systems
Sep 2, 2026Srikumar NayakFederated Learning Federated Averaging
Advanced Artificial Intelligence (AI) Engineering and Cybersecurity Threat Research, LTIMindtree Research, New York City, NY, USA
Scalable and Verifiable Federated Learning for Cross-Institution Financial Fraud Detection
Apr 25, 2026Prajwal Panth, Nishant NigamSecure Aggregation Federated Learning
School of Computer Engineering, KIIT Deemed to be University, Bhubaneswar, India · School of Electronics Engineering, KIIT Deemed to be University, Bhubaneswar, India
Decision-Focused Federated Learning Under Heterogeneous Objectives and Constraints
Apr 21, 2026Konstantinos Ziliaskopoulos, Alexander VinelFederated Learning Heterogeneity
Auburn University

Sep 2, 2026, cs.CRJ/K move · Enter open · S save

Differentially private federated learning with Byzantine-robust aggregation: A cross-domain framework for secure model training in banking and healthcare systems

Srikumar Nayak

Advanced Artificial Intelligence (AI) Engineering and Cybersecurity Threat Research, LTIMindtree Research, New York City, NY, USA

Federated Learning Federated Averaging

Open paper AskPDF arXiv

Federated learning allows banks, hospitals, and other regulated organizations to train a shared model without moving raw records off their own servers, which is attractive wherever data protection law or competitive sensitivity rules out pooling data centrally. Two problems limit how far this promise can be trusted in practice. First, the parameter updates that clients exchange still leak information about local records through gradient inversion and membership inference attacks. Second, an honest averaging rule such as FedAvg has no defense against a subset of clients that submit corrupted or adversarial updates, so a small number of malicious or compromised participants can quietly steer the shared model off course. This paper presents a federated learning framework, DP-BR-FedAvg, that combines a Gaussian-mechanism differential privacy layer with a coordinate-wise trimmed-mean Byzantine-robust aggregation rule, evaluated on a simulated cross-institutional classification task resembling fraud and clinical-risk scoring. Across sixty communication rounds with twenty clients, a quarter of them Byzantine, plain FedAvg collapses on the minority class (F1-score 0.030) while the proposed framework recovers substantially more of the signal (F1-score 0.119) while bounding the privacy loss of any single client's contribution. A Byzantine-robust aggregator with no privacy layer performs best in raw accuracy, quantifying the cost privacy imposes on robustness. The results show that privacy and robustness mechanisms interact rather than simply add, and that system design for regulated, adversarial, cross-institutional settings needs to budget for that interaction.

Abstract

Explore similar work

Differentially private federated learning with Byzantine-robust aggregation: A cross-domain framework for secure model training in banking and healthcare systems

Scalable and Verifiable Federated Learning for Cross-Institution Financial Fraud Detection

Decision-Focused Federated Learning Under Heterogeneous Objectives and Constraints