Cryptocurrency exchanges: Predicting which markets will remain active

Abstract About 99% of cryptocurrency trades occur on organized exchanges with many investors subsequently keeping their digital assets in accounts with cryptocurrency markets. This generates exposure to the risk of exchange closures. We construct a database containing eight key characteristics on 238 cryptocurrency exchanges and employ machine learning techniques to predict whether a cryptocurrency market will remain active or whether it will go out of business. Both in‐sample and out‐of‐sample measures of forecasting performance are computed and ranked for four popular machine learning algorithms. Although all four models produce satisfactory classification accuracy, our best model is a random forest classifier. It reaches accuracy of 90.4% on training data and 86.1% on a test dataset. From the list of predictors, we find that exchange lifetime, transacted volume, and cyber‐security measures such as security audit, cold storage, and bug bounty programs rank high in terms of feature importance across multiple algorithms. On the other hand, whether an exchange has previously experienced a security breach does not rank highly according to its contribution to classification accuracy.