The Effects of Data Imbalance Under a Federated Learning Approach for Credit Risk Forecasting

TL;DR

This work addresses privacy-preserving credit risk forecasting by evaluating Federated Learning (FL) across three models (MLP,LSTM,XGBoost) and three datasets, with particular focus on data imbalance and non-IID distributions. It demonstrates that FL consistently improves non-dominant clients’ performance (average AUC gains of $17.92\%$) and remains robust as the number of clients grows, though gains for dominant clients are less certain. The study extends FL beyond logistic regression, demonstrates performance parity with centralised models in many settings, and highlights practical considerations such as incentives for dominant participants. These findings support FL’s viability for real-world financial collaborations, while underscoring the need for FL-specific hyperparameter tuning and incentive mechanisms to maximize participation and performance.

Abstract

Credit risk forecasting plays a crucial role for commercial banks and other financial institutions in granting loans to customers and minimise the potential loss. However, traditional machine learning methods require the sharing of sensitive client information with an external server to build a global model, potentially posing a risk of security threats and privacy leakage. A newly developed privacy-preserving distributed machine learning technique known as Federated Learning (FL) allows the training of a global model without the necessity of accessing private local data directly. This investigation examined the feasibility of federated learning in credit risk assessment and showed the effects of data imbalance on model performance. Two neural network architectures, Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM), and one tree ensemble architecture, Extreme Gradient Boosting (XGBoost), were explored across three different datasets under various scenarios involving different numbers of clients and data distribution configurations. We demonstrate that federated models consistently outperform local models on non-dominant clients with smaller datasets. This trend is especially pronounced in highly imbalanced data scenarios, yielding a remarkable average improvement of 17.92% in model performance. However, for dominant clients (clients with more data), federated models may not exhibit superior performance, suggesting the need for special incentives for this type of clients to encourage their participation.

Figures (18)

• Figure 1: Overview of federated learning
• Figure 2: Three main categories of federated learning: (a) Horizontal Federated Learning, (b) Vertical Federated Learning and (c) Federated Transfer Learning.
• Figure 3: Overview of XGBoost federated learning.
• Figure 4: MLP performance comparison.
• Figure 5: Performance improvement of MLP with federated learning compared to local models on non-dominant clients.