Cryptocurrency returns forecasting with machine learning

In the recent years, machine learning became significant in many sectors, and finance is no exception. The present diploma thesis is based on exploring the use of machine learning methodologies with LSTM networks, Gated Recurrent Units, and BERT to conduct sentiment analysis which could be utilized for predicting cryptocurrency returns and market sentiment perception to improve decision-making in such a highly volatile cryptocurrency market. It gives a brief history of the cryptocurrency market, followed by describing in detail the methodology regarding data preprocessing: normalization, feature engineering, and creating time series datasets. Development of the LSTM and GRU models for return predictions where BERT is to be used in sentiment analysis from sources such as Reddit. Their respective performances were measured in terms of the Accuracy, Average Daily Return, Average Daily Volatility and T-statistic. Results prove the efficiency of the applied machine learning models both in cryptocurrency return prediction and in gauging sentiment. Trading strategies based on model predictions are backtested with a view to assessing practical profitability. It will also help as a contribution to the application of machine learning in financial forecasting, which will go a long way to help investors and financial analysts understand some capabilities and limitations of those techniques in cryptocurrency markets.