Machine Learning-Based Prediction of Short-Term Bitcoin Volatility Using High-Frequency Data

Using high-frequency Bitcoin transaction and limit order book data, this study investigates the predictability of short-horizon volatility breakouts over a 10-minute horizon, defined as price movements of at least ±0.2%. An XGBoost-based classifier attains an accuracy of 0.767 and a ROC–AUC of 0.794, with breakout precision and recall of 0.634 and 0.446, respectively. SHAP-based interpretability analysis indicates that recent volatility-state measures—most notably Bollinger Band width computed over 20–50 minute horizons—are the dominant predictors. Additional explanatory power arises from trade pressure relative to order book depth and liquidity-fragility proxies derived from limit order book dynamics. Overall, the results suggest that a machine learning model incorporating high-frequency market microstructure information can generate economically meaningful signals for forecasting short-term volatility breakouts in the Bitcoin market.