Prediction of price bubbles in cryptocurrency markets with machine learning methods

This thesis aims to investigate price bubbles in crypto-asset markets—particularly Bitcoin—that can trigger abrupt value declines, and to identify their underlying drivers using machine learning methods. The high volatility and complex market dynamics of cryptocurrencies raise important questions about the existence and detectability of speculative bubbles. In this study, data covering the period from January 1, 2020 to June 5, 2024 are analyzed, and bubble episodes are first identified using the Generalized Supremum Augmented Dickey–Fuller (GSADF) test. In the subsequent stage, these bubble labels are forecast using a range of machine learning and deep learning classification algorithms. For the classification task, the study incorporates not only financial and macroeconomic variables but also engineered technical indicators. Methodologically, the feature engineering phase includes technical measures such as the Relative Strength Index (RSI), Simple Moving Average (SMA), and Average True Range (ATR), alongside macroeconomic indicators. To mitigate the risk of data leakage, all predictors are incorporated with appropriate lags. Model performance is evaluated using metrics well-suited to imbalanced classification, including Test F1 Score, PR-AUC, and ROC-AUC. Empirical results indicate that, relative to traditional machine learning models, the RNN model achieves the strongest performance, attaining a Test F1 score of 0.705 and a PR-AUC of 0.797 under a 70% training–30% testing split. Overall, this thesis proposes a two-stage integrated methodology: (i) econometric bubble detection and labeling via the GSADF test, and (ii) bubble prediction using an RNN-based deep learning architecture.The findings provide tangible contributions in three key areas. First, the proposed framework offers an early-warning mechanism that can help investors mitigate risk during speculative bubble periods. Second, it provides regulators with a practical analytical framework to support market surveillance and oversight. Finally, it contributes to the academic literature by empirically demonstrating that RNN-based approaches—capable of modeling temporal dependence in financial time series—exhibit stronger generalization performance than conventional models.