Background: Like common stocks, Bitcoin price fluctuations are non-stationary and highly noisy. Due to attractiveness of Bitcoin in terms of returns and risk, Bitcoin price prediction is attracting a growing attention from both investors and researchers. Indeed, with the development of machine learning and especially deep learning, forecasting Bitcoin is receiving a particular interest. Methods: We implement and apply deep forward neural network (DFFNN) for the analysis and forecasting Bitcoin high-frequency price data. Importantly, we seek to investigate the effect of standard numerical training algorithms on the accuracy obtained by DFFNN; namely, the conjugate gradient with Powell-Beale restarts, the resilient algorithm, and Levenberg-Marquardt algorithm. The DFFNN was applied to a big dataset composed of 65,535 samples. Results: In terms of root mean of squared errors (RMSEs), the simulation results show that the DFFNN trained with the Levenberg-Marquardt algorithm outperforms DFFNN trained with Powell-Beale restarts algorithm and DFFNN trained with resilient algorithm. In addition, the resilient algorithm is fast which suggests that it could be promising in online training and trading. Conclusions: The DFFNN trained with Levenberg-Marquardt algorithm is effective and easy to implement for Bitcoin high-frequency price data forecasting.
Deep Learning Forecasting in Cryptocurrency High-Frequency Trading
Bekiros S.
2021-01-01
Abstract
Background: Like common stocks, Bitcoin price fluctuations are non-stationary and highly noisy. Due to attractiveness of Bitcoin in terms of returns and risk, Bitcoin price prediction is attracting a growing attention from both investors and researchers. Indeed, with the development of machine learning and especially deep learning, forecasting Bitcoin is receiving a particular interest. Methods: We implement and apply deep forward neural network (DFFNN) for the analysis and forecasting Bitcoin high-frequency price data. Importantly, we seek to investigate the effect of standard numerical training algorithms on the accuracy obtained by DFFNN; namely, the conjugate gradient with Powell-Beale restarts, the resilient algorithm, and Levenberg-Marquardt algorithm. The DFFNN was applied to a big dataset composed of 65,535 samples. Results: In terms of root mean of squared errors (RMSEs), the simulation results show that the DFFNN trained with the Levenberg-Marquardt algorithm outperforms DFFNN trained with Powell-Beale restarts algorithm and DFFNN trained with resilient algorithm. In addition, the resilient algorithm is fast which suggests that it could be promising in online training and trading. Conclusions: The DFFNN trained with Levenberg-Marquardt algorithm is effective and easy to implement for Bitcoin high-frequency price data forecasting.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
