Machine learning models based on bubble analysis for Bitcoin market crash prediction

Abstract

Bitcoin market crashes bring huge economic losses and weaken the global financial system. Thus, predicting Bitcoin market crashes is very important for investors. However, due to the high volatility of Bitcoin prices, accurate prediction of crash events is difficult. This study proposes three components within a hybrid prediction approach to improve Bitcoin market crash prediction for better investment decision-making. First, we offer a hybrid prediction approach that combines various machine learning models with knowledge modeling of bubble phenomenon through the Generalized Supremum Augmented Dickey-Fuller (GSADF) test. Second, we employed the Synthetic Minority Oversampling Technique to address the loss of predictive power due to the severe imbalance between crash and no-crash periods in real-world data. Each machine learning model was trained using an effective oversampling strategy to determine the best predictive performance. Finally, SHapley Additive exPlanations interpretation of outcomes from the predictive model provides practical information for establishing efficient investment strategies. As a result, our crash prediction model based on four long-term bubble cycle information captured via the GSADF test showed significantly improved predictive performance. In particular, the precision-recall area under the curve performance of our prediction model after 7 and 14 days increased by 9.9% and 6.2% compared to original models, which proves that it is effective for mid-to long-term prediction. In addition, factors such as increased interest rates and the continuation of the bubble greatly increased the probability of future crashes in the Bitcoin market. We found that bubble phenomena lasting longer than 14 days significantly increase the probability of a crash.