Using anomaly detection to investigate suspicious activity in sports betting

Abstract

The past few years have seen surges in popularity across the sports betting industry, with several countries removing laws banning sports betting, while betting operators continue to develop their online platforms, enabling customers to place bets from the comfort of their own homes. The research performed involved the application of several different Artificial Intelligence techniques to sports betting data to detect and identify suspicious bets within the dataset, including Anomaly Detection. These suspicious or anomalous bets could represent customers who are particularly desirable for betting operators, customers who are undesirable for betting operators, or customers who are using certain knowledge to attempt to gain an illicit advantage against the betting operator. Suspicious betting patterns could also indicate potential criminal activity, such as fraud, money laundering and match-fixing. The first of two experiments involved the development of an Anomaly Detector that served as a binary classifier, indicating only if a given bet is anomalous or not. Once the Anomaly Detector was implemented, several enhancements were applied to improve the performance of this detector by tackling some of the main challenges known to be present within Anomaly Detection problems. These challenges include the Class Imbalance Problem and Concept Drift. The former is caused since anomalies, by definition, are rarely observed within a dataset. Meanwhile, the latter is caused by changes in the underlying data distribution as time passes. Meanwhile, the second experiment focused on expanding the Anomaly Detector into a multi-class Customer Classifier. This classifier would be capable of differentiating between the various types of anomalies present within the dataset. During the experiment, three different classifiers were compared to establish which would best perform the task at hand. Once the best-performing algorithm was identified, similar improvements to those applied to the Anomaly Detector performing binary classification were applied to the multi-class classifier to address the same two aforementioned problems. From the results obtained, the enhancements that were applied to the Anomaly Detector successfully improved its performance. Furthermore, a Random Forest proved to be the best multi-class classifier, and once enhanced, it proved to be successful when compared to literature.