Intelligent forecasting with machine learning trading systems in chaotic intraday Bitcoin market.

• We employ SVR, GRP, RT, kNN, FFNN, BRNN and RBFNN machine learning models. • We comparatively evaluate AI systems in forecasting high-frequency Bitcoin prices via diverse metrics. • An entropy analysis reveals long memory traits, stochasticity and topological complexity. • The BRNN shows outstanding accuracy while its convergence is remarkably fast. • ANNs efficiently emulate parallel processing in human decision-making under noisy, nonlinear input-output variable spaces. Due to the remarkable boost in cryptocurrency trading on digital blockchain platforms, the utilization of advanced machine learning systems for robust prediction of highly nonlinear and noisy data, gains further popularity by individual and institutional market agents. The purpose of our study is to comparatively evaluate a plethora of Artificial Intelligence systems in forecasting high frequency Bitcoin price series. We employ three different sets of models, i.e., statistical machine learning approaches including support vector regressions (SVR) and Gaussian Poisson regressions (GRP), algorithmic models such as regression trees (RT) and the k -nearest neighbours (kNN) and finally artificial neural network topologies such as feedforward (FFNN), Bayesian regularization (BRNN) and radial basis function networks (RBFNN). To the best of our knowledge, this is the first time an extensive empirical investigation of the comparative predictability of various machine learning models is implemented in high-frequency trading of Bitcoin. The entropy analysis of training and testing samples reveals long memory traits, high levels of stochasticity, and topological complexity. The presence of inherent nonlinear dynamics of Bitcoin time series fully rationalizes the use of advanced machines learning techniques. The optimal parameter values for SVR, GRP and kNN are found via Bayesian optimization. Based on diverse performance metrics, our results show that the BRNN renders an outstanding accuracy in forecasting, while its convergence is unhindered and remarkably fast. The overall superiority of artificial neural networks is due to parallel processing features that efficiently emulate human decision-making in the presence of underlying nonlinear input-output relationships in noisy signal environments. [ABSTRACT FROM AUTHOR]