Modeling and Forecasting the Probability of Crypto-Exchanges' Closure: a Forecast Combination Approach

The popularity of cryptocurrency exchanges has surged rapidly in recent years, with numerous new digital platforms and tokens emerging. However, the issue of credit risk and the reliability of crypto exchanges remains on the agenda, creating a need for some sort of indicator to determine whether a particular exchange is safe to invest through. We examined a unique, hand-collected dataset of 228 exchanges operating from April 2011 to May 2024. Using various machine learning algorithms, we analyzed the key factors leading to the exchange shutdown. Trading volume, exchange lifetime and cybersecutiry scores stood out the most. However, single machine learning algorithms tend to capture different data pecularities and thus make slightly different errors. In this regard, we apply a forecast combination approach by aggregating multiple predictive distributions. Several specifications of generalized linear pool as well as beta-transformed linear pool and beta mixture combinations were investigated. The results showed that the best method is BLP with approximately 4.1\% improvement of forecast accuracy based on a robust H-measure which handles the problem of misclassification in the presence of imbalanced classes.