Bitcoin time-series volatility forecasta comparative analysis of the performance of machine learning algorithms

Resumen

This Thesis undertakes a comprehensive examination of Bitcoin volatility, investigating its connections with economic, financial, technological variables and various market dynamics. Multivariate statistical time series models (GARCH family) and Artificial Intelligence techniques (supervised and unsupervised learning) are employed to unravel these intricate relationships. The Conditional Correlation approach, in particular the CCC-MGARCH model, emerges as a frontrunner in modelling Bitcoin volatility, taking into account the variables considered. Among the machine learning models, Ridge Regression stands out, outperforming Ensemble and K-Nearest Neighbors in predicting Bitcoin volatility. This finding underlines the effectiveness of internal regularisation in capturing complex interdependencies. The study is in line with previous research, which suggests the accuracy of Random Forest and Ridge Regression in predicting Bitcoin volatility. The unsupervised learning algorithms DBSCAN and HDBSCAN do not perform optimally due to the inherent instability of Bitcoin data. The K-Means analysis, however, reveals valuable data, such as volatility declines around the exponential moving average (EMA) and correlations between the relative strength indices (RSI) of the selected assets. In conclusion, supervised learning models excel in predicting Bitcoin volatility, outperforming unsupervised clustering approaches in the complex financial landscape. However, some limitations remain, such as the difficulty of encompassing all influencing factors and the inherent unpredictability of Bitcoin volatility. The study advocates exploring additional variables associated with news, market sentiment and external events to refine predictions, recognising that Bitcoin volatility remains a difficult task.