Detectability of Sports Betting Anomalies Using Deep Learning-based ResNet: Utilization of K-League Data in South Korea

Sports match-fixing refers to the act of pre-determining the results of a game. Match-fixing fundamentally undermines competition in sports, and it harms society as a whole. Match-fixing has large as it is conducted as a secret transaction. Consequently, finding evidence regarding this illegal activity or detecting it is difficult. Therefore, a system should be built to detect match-fixing to prevent it.

This study aims to detect betting anomalies in sports events through dividend graphs.

We collected the odds data for the K-League from 2010 to 2020 and converted the data into graph images to generate 3101 graph images. The collected data was analyzed using ResNet to classify them into normal games (2,464 games) and abnormal games (637 games) based on an image classification method. The ResNet model was trained for 100 epochs, and as a result, values below 0.05 were derived as the loss values of the training and test data, respectively.

After performing the validation with the test data of 50 normal and 27 abnormal games, it was found that the accuracy in deriving normal games was 90%. Furthermore, match-fixing games were derived with an accuracy of 74.1%. Therefore, the model was accurate for 65 out of 77 games, showing that the model’s accuracy was 84.44%.

The results demonstrate the model’s value as a method for detecting sports match-fixing. Additionally, it can aid in eradicating sports match-fixing by providing the basic data for undertaking detailed match-fixing investigations.