Predicting cryptocurrency volatility: The power of model clustering

This study examines the predictability of cryptocurrency volatility, a critical challenge given the extreme fluctuations characteristic of these assets. Existing literature highlights the limitations of single-model approaches in predicting such volatility. Using high-frequency data from Binance for ten cryptocurrencies spanning diverse market capitalizations, we systematically evaluate various forecast combination approaches. Our analysis compares traditional linear heterogeneous autoregressive and nonlinear realized volatility models with advanced forecast combination techniques. Results indicate that the winning combination approach significantly improves predictive accuracy over individual models and alternative combination techniques. This enhanced performance arises from its ability to leverage latent groupings among forecasting model weights effectively. Furthermore, we demonstrate the economic value of these improved forecasts, quantifying an average utility gain equivalent to 3.46% of wealth for risk-targeting investors. These findings provide novel insights into volatility forecasting and suggest practical implications for investors seeking to optimize risk management strategies in cryptocurrency markets.