Complexity analysis and forecasting of variations in cryptocurrency trading volume with support vector regression tuned by Bayesian optimization under different kernels: An empirical comparison from a large dataset

Abstract

When cryptocurrency markets generate billions of dollars, it becomes interesting to forecast variation in volume of transactions for better trading and for better management of blockchain platforms. This study investigates how kernel choice influences the forecasting performance of the support vector regression (SVR) in predicting cryptocurrency trading volume. Three common kernels are considered; namely, linear, polynomial, and radial basis function (RBF). In addition, we make use of Bayesian optimization (BO) method to tune key parameters of the SVR, hereafter referred as SVR-BO. Besides, we examine the nonlinear dynamics of variation in volume of transactions by computing Hurst exponent, sample entropy, and largest Lyapunov exponent and found evidence of anti-persistence, significant randomness, and presence of chaos. Well-known ARIMA process, Lasso regression and Gaussian regression are used as benchmark models in the forecasting task. The root mean of squared errors (RMSE) and mean average error (MAE) are adopted as main performance metrics. Forecasting simulations are applied to thirty cryptocurrencies. The results from 180 experiments show that the SVR-BO with RBF kernel outperforms all models when used to predict next-day trading volume while SVR-BO with polynomial kernel outperforms all remaining models when used to predict next-week trading volume. Besides, Gaussian regression performs better than ARIMA process and Lasso regression on both daily and weekly data.