Real-Time Pump and Dump event detection for Cryptocurrencies

During the last years the cryptocurrency field has gained a lot of momentum, becoming more and more prominent and mainstream. Parallel to this wave of excitement, many new projects have sprung up, together with a great amount of frauds, which can easily proliferate due to the lack of regulations and controls. Many inexperienced investors have been tricked into fraudulent schemes, with the false promise of high returns. One of the most popular frauds is the so-called "pump and dump" scheme, where an organized group of people wittingly manipulates the price of an asset by generating sudden large buy volumes which drive up the price. Inexperienced investors end up losing their capital in this operation. This thesis work studies this widespread phenomenon and proposes an algorithmic system based on artificial intelligence technologies to detect pump and dumps in real-time. This task is really challenging, as pump and dump events can be very rapid and in order to have an effective detection, the detector must be able to trigger an alert within the first seconds of market manipulation. A few other detectors are already present in the literature. These systems achieve great performance, but their main shortcoming is speed, which can lead to late detections. The purpose of this work is therefore to improve the detection speed: the proposed system can individuate pump and dumps right when they start (in the first second of market manipulation). In addition, a novel algorithmic labeling strategy for pump and dump datasets is proposed, with the aim of both facilitating the work of the human annotator and, also, making the dataset creation process scalable.