Study and development of machine learning-based cryptocurrency trading systems

Since the release of Bitcoin, cryptocurrencies have gained more and more attention, becoming an important financial reality. Market capitalization exploded when the Bitcoin/USD pair reached its all-time high in December 2017, attracting investors from retail, professional and institutional markets in a novel gold run. Existing studies on Decision Support Systems (DSS) and Automated Trading Systems based show pertinence of such techniques to traditional markets, while their application to cryptocurrencies is still a study subject. This work proposes the analysis of multiple cryptocurrencies among the most popular by market capitalization between 01/2011 and 01/2019, combining daily market data with a selection of technical analysis indicators and blockchain-derived metrics to build and analyse different trading systems based on popular machine learning algorithms and ensemble methods both in terms of performance and relevant feature interactions. Results show how prediction outcomes are generally following trends, with model precision ranging between 40% and 60%. However, when analysed through metrics who take input bias into account such as index-based accuracy (IBA), very few models reach the skill threshold, implicating class imbalance in the training data affects classification results. Trading simulation shows how the proposed systems are profitable in both bear and bullish markets yet fail to identify patterns leading to high volatility events characterising the cryptocurrency markets, giving the baseline strategy a lead over longer timespans. The work also explores the reasons behind machine learning algorithms' decisions. It applies a state-of-the-art explainable model, namely SHAP, to highlight the features that mostly influence the performance of cryptocurrency price forecasting.