Wash trade is a technique to create the illusion that an instrument has high demand in the stock market. Brokers can unfairly profit from this asking for inflated commission fees. The existing research in wash trade detection usually assumes that every transaction in the stock market is known. However, several intermediaries in the business can only capture the actions of their direct clients, and they have to flag and analyze suspicious limit orders to comply with regulatory rules. We propose adopting anomaly detection to find wash trade transactions using partial information. First, we present a novel user behavior representation based on strings. Then we experiment with string similarity measures and indexing techniques to cluster accounts efficiently. For every cluster, we train Neural Turing Machines that can predict the next action of a user given his event history. Finally, we calculate an aggregated anomaly score based on the distribution of events in every model. The project also focus in the practical implementation of the algorithm under a streaming environment, covering issues such as concept drift and analyzing Big Data technologies.