Stock price manipulation detection using a computational neural network model

We investigated the characteristics of stock price manipulation. Two manipulation models were studied: pump-and-dump and spoof trading. Pump-and-dump is a procedure to buy a stock and push its price up. Then, the manipulator dumps all of the stock he holds to make a profit. Spoof trading is a procedure to trick other investors that a stock should be bought or sold at the manipulated price. We constructed mathematical models that use level 2 data for both procedures, and used them to generate a training set consisting of buy/sell orders within on order book of 10 depths. Order cancellations, which are important indicators for price manipulation, are also visible in our level 2 data. In this paper, we consider a challenging scenario where we attempt to use less-detailed level 1 data to detect manipulations even though using level 2 data is more accurate. We implemented feedforward neural network models that have level 1 data, containing less-detailed information (no information about order cancellation), but is more accessible to investors as input. The neural network model achieved 88.28% for detecting pump-and-dump but it failed to model spoof trading effectively.