Identification of Scams in Initial Coin Offerings With Machine Learning.

Abstract

Following the emergence of cryptocurrencies, the field of digital assets experienced a sudden explosion of interest among institutional investors. However, regarding ICOs, there were a lot of scams involving the disappearance of firms after they had collected significant amounts of funds. We study how well one can predict if an offering will turn out to be a scam, doing so based on the characteristics known ex-ante. We therefore examine which of these characteristics are the most important predictors of a scam, and how they influence the probability of a scam. We use detailed data with 160 features from about 300 ICOs that took place before March 2018 and succeeded in raising most of their required capital. Various machine learning algorithms are applied together with novel XAI tools in order to identify the most important predictors of an offering’s failure and understand the shape of relationships. It turns out that based on the features known ex-ante, one can predict a scam with an accuracy of about 65–70%, and that nonlinear machine learning models perform better than traditional logistic regression and its regularized extensions.