Cryptocurrencies: Trading and Return Prediction via Machine Learning Methods

Abstract

In this thesis we study to which extent modern machine learning algorithms and classical time series methods are capable of predicting the one-day-ahead price trend of the well known Bitcoin cryptocurrency. More specifically, we address this forecasting task by implementing four machine learning classifiers. These include the logistic regression model and three neural network architectures, namely the fully-connected-, GRU-, and convolutional neural network. We compare their results with the prediction performances of the classical ARMA model and the trend following momentum strategy. Taking into account that an algorithm’s forecasting accuracy inevitably depends on the quality of training data, we analyse a carefully selected set of 25 features. These include blockchain related market forces of supply and demand, global macroeconomic and financial development indicators and competing cryptocurrency market price data. Furthermore, we utilise attractiveness measures like Google Trends and Wikipedia Pageviews and additionally use natural language processing to create a Google News feature, which ideally captures the market sentiment concerning Bitcoin. We find that the best GRU setup achieves a remarkable test accuracy of 62% in predicting the next-day Bitcoin price trend, followed by the convolutional neural network and logistic regression, which attain 59% and 57%, respectively. With a test accuracy of only 52%, the fully-connected neural network and the ARMA model rank last not being able to surpass the simple strategy of going long in Bitcoin. However, all these results are excelled by the computationally cheap and intriguingly simple momentum strategy. It equalises the GRU network’s test accuracy of 62%, but achieves better performance with regards to its annualised Sharpe ratio and Jensen’s alpha.