A Transformer Model for Stock Price Manipulation Detection in the Stock Exchange of Thailand

Machine learning is effective in detecting stock manipulation because it can extract key characteristics from wide variety of manipulation tactics. However, to keep up with sophisticated manipulators who constantly alter their strategies to avoid getting caught, the model and data must be updated on a regular basis. This paper explored the effectiveness of the state-of-the-art Transformer model to detect stock price manipulation. We used data from the Stock Exchange of Thailand to train the model in a supervised manner. We synthesized various degrees of pump-and-dump patterns and injected them into the normal dataset for training. The test set consisted of actual SEC-reported manipulation cases from 2004 to 2016. The model can detect five out of six real manipulation cases prosecuted by the SEC with a low false-positive rate.