We analyze the quality of Bitcoin volatility forecasting of GARCH-type models applying different volatility proxies and loss functions. We construct model confidence sets and find them to be systematically smaller for asymmetric loss functions and a jump robust proxy