Modeling cryptocurrency failure using deep learning approaches and a post-hoc interpretability algorithm

Abstract

The lack of enforced trading termination in the cryptocurrency market allows for an exploration of the “natural death” of tradable assets. We propose two deep learning survival models to study this phenomenon and a post-hoc interpretability algorithm to interpret the results and test the hypothesis. The proposed deep learning survival models outperform the time-dependent Cox regression in both prediction performance and interpretation flexibility. Our results indicate that lower trading volume, market capitalization, social media attention, and higher price volatility predict increased cryptocurrency “death” hazards.