Graph Neural Networks and Federated Learning for Privacy-Preserving Volatility Prediction in Global Multi-Asset Financial Markets

Multi-asset volatility forecasting in global financial markets allows risk management, portfolio allocation, and oversight processes to function appropriately. In contrast, traditional approaches in econometrics or conventional ML methods would struggle to model the complex, dynamic, spatial-temporal dependencies existing across heterogeneous financial assets. Strict data privacy norms obstruct collaborative modeling among financial institutions. This work, therefore, aims to develop a novel, unified framework that symbiotically combines DT-GNNs and an AFL paradigm to enable robust, scalable, and privacy-preserving volatility prediction. The proposed DT-GNN architecture is modeled to dynamically analyze evolving inter-asset relations from temporally weighted graph convolutions coupled with attention mechanisms to grasp the fine-grained cross-sectional and temporal structures in the market. Concurrently, the AFL protocol facilitates decentralized training across heterogenous data silos by adaptively weighting client contributions based on local model performance, thus accelerating convergence and enhancing predictive accuracy while safeguarding proprietary data. Extensive empirical analyses conducted on a diverse dataset comprising equities, foreign exchange, cryptocurrencies, and commodities demonstrate that our approach significantly outperforms leading benchmarks—including GARCH variants, LSTM, Transformer models, and federated GNN baselines—across key metrics such as root mean squared error (RMSE), mean absolute error (MAE), and volatility spike detection accuracy. Complementary interpretability assessments via SHAP elucidate feature importance and bolster the model’s explanatory power, an essential attribute for adoption in regulated financial environments. Our results establish the proposed DT-GNN + AFL framework as a state-of-the-art, practical solution for multi-asset volatility forecasting that effectively balances accuracy, scalability, interpretability, and stringent privacy requirements, thereby advancing collaborative financial analytics in complex, data-sensitive ecosystems.

کلیدواژه ها:

Multi-asset volatility forecasting ، Dynamic Temporal Graph Neural Networks (DT-GNN) ، Adaptive Federated Learning (AFL) ، privacy-preserving machine learning ، financial market risk modeling