We estimate the expected shortfall of four major cryptocurrencies using various error distributions and GARCH-type models for conditional variance. Our aim is to examine what distributions perform better and to check what component of the specification plays a more important role in the estimation of the expected shortfall. We evaluate the performance of the estimations with a rolling-window backtesting technique, based on the multinomial test proposed in Kratz, Lok and McNeil (Journal of Banking and Finance, 2018). Our results highlight the importance of estimating the expected shortfall of Bitcoin using a generalized GARCH model and a non-normal error distribution with at least two parameters; however, a GARCH model and a Student-t error distribution performs well with other cryptocurrencies. Further, in order to allow for the possibility of structural change in the return series we also present a model with jumps in the econometric specification.