Abstract
The advances of blockchain technology and its global recognition over the last years enabled the rise of digital currencies, also known as cryptocurrencies. These currencies are characterized by unique properties like pseudonymity, low trading fees, and minimal barriers of entry which made them increasingly interesting as investment opportunities.Additionally, these currencies are considered highly volatile and the efficient market hypothesis does currently not hold true according to researchers, making them the ideal target for automated analysis and trading.At the same time, deep neural networks and novel neural-network architectures have been producing promising research results for time-series predictions and sentiment-analyses.However, research on the combination of deep-learning, technical indicators, and financialsentiment analysis in the field of cryptocurrency market predictions is still scarce. Theaim of this thesis is to explore this research gap and provide answers to the complexquestions associated with it.To reach that goal we developed and evaluated multiple deep neural networks to generate trading strategies for the Bitcoin cryptocurrency. Specifically, we focused on the optimization of the structure and hyper parameters of the neural networks, explored the space of risk adjustable target values, tested alternative input sources, and developed a simulation engine for the generated trading strategies.The findings of our experiments confirm the hypothesis that cryptocurrencies open vastopportunities for profitable automated trading. Our experiments show that AI-basedtrading can significantly improve profitability compared to a buy-and-hold strategy while simultaneously reducing the risk associated with it.
Reference
Muhm, T. (2021). Cryptocurrencies: Deep learning for sentiment & market analysis [Diploma Thesis, Technische Universität Wien]. reposiTUm. https://doi.org/10.34726/hss.2021.71386