Evaluating Predictive Robustness of Machine Learning Models During Black Swan Crises: Insights from Bitcoin Price Forecasting

Predicting Bitcoin prices has always been challenging due to its high volatility and lack of linearity, and this challenge becomes even more pronounced in the case of market disruptions. In this context, this paper investigates the appropriateness of four machine learning models in forecasting, namely XGBoost, Long Short-Term Memory, Bagging Ensemble, and Stacking Ensemble, during two recent Black Swan periods, such as the COVID-19 pandemic and the Russia–Ukraine war. For this purpose, a dataset of 1240 Bitcoin daily closing prices from February 23, 2020, to August 8, 2023, was considered for the prediction purpose. The next day, Bitcoin prices were forecasted, and the prediction accuracy was measured against root mean square error and mean absolute percentage error. The results revealed that the Bagging and Stacking Ensemble was the most precise model during both Black Swan events. In contrast, Long Short-Term Memory (LSTM) and XGBoost were the most and least accurate, respectively. The finding indicates the robustness of ensemble-based approaches to coping with financial uncertainty. The study is instrumental because it allows comparing multiple models during extreme conditions, which provides vital insights for traders, analysts, and policymakers striving to identify the most accurate sources. Finally, this study contributes to the limited body of AI-driven financial forecasting literature focusing on models during actual Black Swan events instead of hypothetical situations.