Automatisierte Prognose der Entwicklung von Kryptowährungspreisen

Since the introduction of Bitcoin, cryptocurrencies have become very attractive as an alternative digital payment method and a highly speculative investment. With the rise in computational power and the growth of available data, the artificial intelligence concept of deep neural networks had a surge of popularity over the last years as well. With the introduction of the long short-term memory (LSTM) architecture, neural networks became more efficient in understanding long-term dependencies in data such as time series. In this thesis, we combine these two topics, by using neural networks to make a prognosis of cryptocurrency prices. In particular, we test if LSTM based neural networks can produce profitable trading signals for the cryptocurrency Ethereum. We experiment with different preprocessing techniques and different targets, both for price regression and trading signal classification. We evaluate two LSTM based networks and one convolutional neural network (CNN) LSTM hybrid. As data for training we use historical Ethereum price data in one-minute intervals from August 2017 to December 2018. We measure the performance of the models via backtesting, where we simulate trading on historic data not used for training based on the models predictions. We analyze that performance and compare it with the buy and hold strategy. The simulation is carried out on bullish, bearish and stagnating time periods. In the evaluation, we find the best performing target and pinpoint two preprocessing combinations that are most suitable for this task. We conclude that the CNN LSTM hybrid is capable of profitably forecasting trading signals for Ethereum, outperforming the buy and hold strategy by roughly 30%, while the performance of the other two models was disappointing.