Limelight: Real-Time Detection of Pump-and-Dump Events on Cryptocurrency Exchanges Using Deep Learning

Abstract

Following the birth of cryptocurrencies back in 2008, internet investment platforms called exchanges were created to constellate these cryptocurrencies. Allowing investors to sell and buy assets equitable and agile over a single interface. Exchanges now have become popular and carry out over 99% of all daily transactions, totaling hundreds of millions of dollars. Despite that exchanges handling enormous quantities of money, the industry remains mostly unregulated.

As long as these exchanges remain unregulated, they are and will continue to be susceptible to price manipulation schemes since they are legal to perform by law. Over the years, exchanges have grown into an attractive field where scammers execute various frauds that aims to leech assets from ordinary investors. One particular scheme has risen in popularity over the years and often observed at exchanges, and that is pump-and-dump. This scheme has a history from all the way back in 1700 and is still active and troublesome for investors today.

In this thesis, we present Limelight, a system that seeks to detect pump-and-dump in real-time using deep learning. Throughout this thesis, we retrieved, prepared, labeled, and processed a dataset to train a model that identifies pump-and-dumps. With high accuracy, the model surpasses previously proposed models in the detection of pump-and-dumps.