This paper proposes the use of threshold heteroskedastic models which integrate threshold nonlinearity [Tong, H (1978). On a Threshold Model, pp. 575–586. Netherlands: Sijthoff & Noordhoff; Tong, H and KS Lim (1980). Threshold autoregression, limit cycles and cyclical data. Journal of the Royal Statistical Society. Series B (Methodological), 3, 245–292.] and GARCH-type conditional variance for modeling Bitcoin returns to provide an understanding on the huge volatility that Bitcoin has been famous for. Specifically, the model attempts to identify different regimes throughout the history of Bitcoin using the different available Bitcoin network characteristics, such as cost per transaction, number of transactions per block, number of active addresses and number of transactions. Estimation and diagnostic checks are performed using Markov chain Monte Carlo methods. In the empirical analysis, we show that our model is able to identify periods of crashes as one of these regimes, which is also a period of declining returns and declining number of active users. We also find that the number of users and the number of transactions determine the magnitude or persistence of a crash period.