Predictive Accuracy of Bitcoin Price Movements: A Comparative Analysis of Machine Learning Models and ARIMA

Abstract

This paper focuses on assessing the accuracy of predicting the direction of Bitcoin's price in
USD. Historical price data is extracted from the Bitcoin Price Index, and the task is approached
with varying levels of success by leveraging Bayesian-optimized Support Vector Regression
(SVR) methods and Long Short-Term Memory (LSTM) networks. Among the methods
employed, LSTM achieves the highest classification accuracy at 52% and a Root Mean Square
Error (RMSE) of 8%. In addition, the popular ARIMA model for time series forecasting is
incorporated for comparison with machine learning models. As anticipated, the non-linear
machine learning approaches outperform ARIMA, which exhibits poor performance. Lastly, the
study includes a benchmarking analysis of both machine learning models implemented on both a
GPU and a CPU, revealing a 67.7% improvement in training time for the GPU-based
implementation.