Not logged in.
Quick Search - Contribution
Contribution Details
| Type | Master's Thesis |
| Scope | Discipline-based scholarship |
| Title | Forecasting Bitcoin Volatility with Adaptive Machine Learning |
| Organization Unit | |
| Authors |
|
| Supervisors |
|
| Language |
|
| Institution | University of Zurich |
| Faculty | Faculty of Business, Economics and Informatics |
| Number of Pages | 63 |
| Date | 2022 |
| Abstract Text | Bitcoin was the overall best-performing asset in the last decade and the popularity of the most successful cryptocurrency seems unbroken with an average daily trading volume of more than 100 million US dollars since 2020. Despite the impressive past returns, the high volatility is widely seen as one of the most deterring factors for investors. Thus, for risk management and to optimise investor’s decision-making the analysis of volatility is crucial. This thesis uses adaptive machine learning models to forecast next day’s directional volatility movement with 21 features from five different categories: (1) two Bitcoin price features, (2) four technical indicator features, (3) four economic features, (4) six Bitcoin metric features, and (5) five financial market features. The primary objective is building a binary classification problem to predict an increase or decrease in daily volatility with boosting algorithms. Both the Adaptive Boosting classifier (AdaBoost) as well as the Extreme Gradient Boosting classifier (XGBoost) show promising results. XGBoost achieved the highest mean accuracy with 59.6%. The feature importance analysis suggests the advantage of selecting features from different categories. Furthermore, a multi-category model is built to separate the 10% biggest positive and negative volatility changes from minor increases or decreases. This model achieved a mean accuracy of 46.6%. The main conclusion of the analysis is that boosting algorithms show promising results for binary Bitcoin volatility forecasting but parameter tuning and potential overfitting are limits to the predictive power of the models. However, the results still indicate that a lean built machine learning model based on 21 features can make limited predictions of short-term Bitcoin volatility. This can be useful in terms of risk management and investment decisions. |
| Export |
BibTeX
|