A1500
Title: Forecasting Bitcoin volatility with non-Gaussian realized stochastic volatility models
Authors: Yosuke Mori - Keio University (Japan) [presenting]
Makoto Nakakita - RIKEN (Japan)
Teruo Nakatsuma - Keio University (Japan)
Abstract: An empirical comparison of Gaussian and non-Gaussian realized stochastic volatility models for Bitcoin is presented. Using daily Bitcoin returns and one-minute realized volatility from Binance (2022-2024), the analysis compares seven return-error specifications within a common realized stochastic volatility architecture: Gaussian, variance gamma, skew variance gamma, Student t, skew Student t, Laplace, and asymmetric Laplace. The non-Gaussian models are estimated through a unified generalized inverse Gaussian latent-scale representation, so the comparison isolates the empirical role of heavy tails and asymmetry while keeping the state and the realized volatility measurement equation fixed. Forecast evaluation uses a 730day rolling design, the two-scale realized volatility proxy, mean squared error, quasi-likelihood, Giacomini and White tests, ninety-five percent predictive-interval diagnostics, and the widely applicable information criterion. Model ranking is criterion dependent rather than winner take all. The skew Student t model attains the best widely applicable information criterion, the Student t model attains the best mean squared error, the asymmetric Laplace model attains the best quasi-likelihood loss, and the skew variance gamma model yields the narrowest near nominal predictive interval. The results show that non-Gaussian return laws materially change model ranking for Bitcoin across density fit, point forecasting, and interval calibration.
