Essays in cryptocurrencies’ forecasting and trading with technical analysis and advanced machine learning methods

This thesis mainly emphasizes two prediction fields in the cryptocurrency market: factor analysis and model examination. The first section summarises the general introduction, theoretical background, and description of performance metrics used in the empirical study (Chapter 3-5) are summarized in the first section (Chapter 1-2). Then, in Chapters 3 and 4, technical analysis and fundamental factors combined with statistical models are employed to explore the forecasting ability and profitability in the cryptocurrency market. Finally, in Chapter 5, advanced machine learning algorithms combined with leverage trading strategies and narrative sentiments are used to predict the Bitcoin (BTC) market. Chapter 3 examines technical analysis’s profitability and predictive power on cryptocurrency markets. This Chapter adopts the universe of technical rules proposed by Sullivan et al. (1999), while for data snooping purposes, I apply the Lucky Factors (LF) method proposed by Harvey and Liu (2021). Six mainstream cryptocurrencies and one cryptocurrency index from 2013 to 2018 are examined. The results demonstrate that short-term signals generated by technical rules outperform the traditional buy-and-hold strategy. However, the LF methodology shows that none of the top-performing rules in terms of profitability is consistent with actual forecasting performance. The purpose of Chapter 4 is to investigate the prediction of cryptocurrency returns by applying a large pool of factors from both technical and fundamental aspects. The results find that most trading rules perform better than the buy-and-hold strategy, especially the moving average rules. However, this profitability may not be genuine but comes from data-snooping bias. In this way, a larger pool of factors from several aspects, including blockchain information, technical indicators, online sentiment indices, and conventional financial and economic indicators, is implemented from 08/08/2015 to 08/12/2018. The overall results suggest the new proposed technical indicator, Log-price Moving Average (PMA) ratio, a moving-average likely ratio has significant forecasting ability in cryptocurrencies after taking data-snooping bias into account. Chapter 5 explores the forecasting ability of machine learning (ML) algorithms in the BTC market by combining the narrative sentiments and leverage trading strategy. First, the forecasting framework starts by selecting a pool of individual models. Secondly, ML algorithms are used further to improve the predictive performance of the individual model pool. Thirdly, both the best single predictor and ML models are fed into the process of forecasting ability examination, constructed by three different metrics. This step also takes data-snooping bias into account. At last, leverage trading strategies combined with narrative sentiments are applied to all forecasting models to examine their profitability. The results suggest that ML models consistently outperform the best individual model in forecasting ability and profitability. Gradient Boost Decision Tree (GBDT)-the family has the best performance