Federated Learning with Efficient Local Adaptation for Realized Volatility Prediction
Abstract: Financial markets are inherently complex, with private trading activities distributed across various exchanges and platforms, leading to isolated datasets and fragmented data sources. Learning from limited local data leads to inaccurate realized volatility prediction due to incomplete representations of market dynamics. Federated learning (FL) can foster collaborative insights while ensuring privacy and regulatory compliance across diverse trading platforms. However, heterogeneity in datasets and dynamic participation making FL for financial markets slow and inaccurate. To address this issue, we propose Federated Learning with Efficient Local Adaptation (FLELA). The key idea is to enhance the local model with probabilistic techniques, including local linearization of the global model and a crucial optimization step to fine-tune parameters, so each participants can apply enhanced local model to achieve higher accuracy. Through extensive experimental evaluations, FLELA consistently outperforms existing federated learning algorithms, demonstrating superior predictive accuracy and efficiency in realizing volatility prediction. Even in the face of significant data fragmentation across massive trading venues, the proposed FLELA can achieve mean loss of $7.358 \times 10^{-5}$, VaR95\% of $2.284 \times 10^{-4}$, and CVaR95\% of $3.978 \times 10^{-4}$ in merely five rounds of FL, which is one order better than the state-of-the-art FL approaches, underscoring its efficacy and superiority.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Zhiyu_Zhang1
Submission Number: 3393
Loading