Cross-Sector Risk Forecasting Using Federated Machine Learning: Applications in Finance, Healthcare and Urban Planning
Keywords:
Federated Learning, Personalized FedProx, Cross-Sector Learning, Risk Forecasting, Financial Risk Prediction, Healthcare Analytics, Urban Planning, Privacy-Preserving Machine Learning.Abstract
Federated learning has emerged as a potential paradigm for collaborative intelligence because it allows many businesses to train machine learning models together without sharing raw data. However, traditional federated learning algorithms frequently assume homogenous data distributions and shared feature spaces, which limits their application to diverse cross-domain contexts. This study introduces a Personalized Federated Proximal (FedProx) architecture for collaborative risk forecasting in three distinct sectors: finance, healthcare, and urban planning. The suggested architecture includes sector-specific input adapters, a shared residual backbone, private residual adapters, and task-specific prediction heads to strike a compromise between collaborative knowledge sharing and domain-specific expertise. Adaptive sector-weighted aggregation is used to improve global model optimization while maintaining local characteristics. The framework is tested on publicly available datasets using a rigorous process that includes five random seeds and three rolling temporal folds for urban forecasting, totaling 15 repeated trials. Performance is measured using Accuracy, Balanced Accuracy, F1-score, Matthews Correlation Coefficient (MCC), ROC-AUC, and PR-AUC. Experimental results reveal that the proposed framework performs similarly to the Sector FedAvg baseline in financial risk forecasting while consistently improving urban forecasting, particularly in ROC-AUC and PR-AUC. Although healthcare remains a difficult prediction problem due to the complexity of the leakage-safe composite target, the suggested strategy improves overall classification accuracy while retaining consistent performance across multiple experiments. These findings show that personalized federated learning is a viable method for collaborative risk prediction across diverse application domains while limiting negative transfer via sector-specific adaptation.
