Forecasting Bitcoin Implied Volatility Changes with On-Chain, Macro,and Sentiment Signals: Benchmarking XGBoost against GARCH Models

Predicting changes in Bitcoin's Implied Volatility (IV) is crucial for derivatives pricing and risk management but presents significant challenges for traditional time-series models. We propose an interdisciplinary machine learning approach using XGBoost trained on a feature set combining on-chain metrics (e. g. , SOPR), macroeconomic indicators (e. g. , Fed Rate momentum), social sentiment, and technical features including Continuous Wavelet Transforms (CWT). Our model demonstrates statistically significant predictive power for the 1-day change in the DVOL index, achieving an out-of-sample R2 of 0. 0527, markedly outperforming a baseline GARCH (1, 1) model (R2 = -0. 0107). SHAP analysis identifies key predictors including recent price changes (pricechange₁d), on-chain sentiment (sopr), and macroeconomic momentum (FedRateₘom). However, translating this statistical edge into economic profit proved challenging with a simplified backtest methodology. A strategy trading based on the predicted IV change direction, incorporating risk controls (confidence filtering, 5% stop-loss resulting in a Max Drawdown of -1. 18%), yielded a Sharpe Ratio of -13. 40, underperforming the Buy & Hold benchmark (Sharpe 1. 78). We attribute this negative economic result primarily to the PnL normalization proxy used (dividing IV point changes by spot price), which significantly dilutes the signal relative to transaction costs. As discussed, alternative PnL calculations reflecting realistic options Vega scaling suggest potential for profitability. Our findings confirm a predictive signal exists but highlight the critical gap between statistical significance and capturing alpha economically, emphasizing the need for strategy refinement and realistic PnL modeling in future work.