The modern financial electronic exchanges are dynamic and fast-moving marketplaces where billions of dollars are exchanged daily. However, these exchanges are also susceptible to manipulation and fraudulent activities. Traditionally, identifying such activities has been a task assigned to humans. However, there has been a recent shift towards automating these processes using machine learning and artificial intelligence.

Fraud detection in financial exchanges is closely related to anomaly detection, which usually involves unsupervised learning techniques due to the limited availability of labeled data. Although there is a scarcity of labeled data, a small amount of such data does exist. This research article aims to assess the effectiveness of a deep semi-supervised anomaly detection technique called Deep SAD in detecting fraud in high-frequency financial data. The evaluation is conducted using exclusive proprietary limit order book data from the TMX exchange in Montréal, along with a small set of labeled instances of fraud. Deep SAD is compared to its unsupervised predecessor, and the results demonstrate that incorporating a small amount of labeled data can significantly enhance the accuracy of an unsupervised anomaly detection framework.