Federated Learning Approach for Bitcoin Price Prediction

Abstract

The intricate and unpredictable nature of cryptocurrency markets has brought Bitcoin price prediction into the spotlight because of the relatively high volatility levels of cryptocurrencies. Conventional centralized machine learning solutions pose challenges on the issue of data privacy, security, and scalability, especially with financial applications. To respond to these issues, this paper outlines an overall execution of a federated learning system to predict Bitcoin prices. K-Nearest Neighbours, Decision Tree, Linear Regression, and Federated Long Short-Term Memory (FL-LSTM) model are federated, trained, and tested on past Bitcoin market data. Within the proposed framework, the process of model training is executed at each of several clients locally, and only model parameters or predictions are transmitted without data privacy. Experimental data reveal that classical federated machine learning models have poor performance in modelling complex price dynamics. Although Federated Linear Regression reflects similar goodness of-fit, the FL-LSTM proposed model is always associated with the lower prediction error and is highly close to the real price movements. Also, the FL-LSTM model is used to predict the short-term future, which proves that this model can be useful to anticipate future fluctuations in Bitcoin prices. The results affirm that federated learning, which has been combined with deep learning models, is a viable and privacypreserving solution in cryptocurrency prediction in a decentralized setting. Root Mean Square Error (RMSE) and$\mathbf{R}^{\mathbf{2}}$were used to measure the proposed models. The results of the experiment show that the FL-LSTM model possesses the lowest prediction error, and it is much closer to real Bitcoin price trends than the other federated models.