Abstract

Recently, the blockchain based cryptocurrencies and crypto tokens have started to attract significant interest. Crypto tokens that are sold on existing blockchains such as Ethereum have been used to raise significant funding for many start-ups. At the same time, many crypto tokens have failed and resulted in significant financial loss for their investors. This raises an important question: Can we predict the anomalous crypto tokens using the transaction graph data stored on the blockchain? Unfortunately, due to dynamic and sparse nature of the crypto token transaction graphs, existing graph analysis techniques are not directly applicable. Instead, we propose novel techniques based on topological data analysis and functional data depth that allow us to extract features that are useful for anomaly prediction. Our extensive empirical analysis show that the proposed techniques significantly outperform baseline models.