Predicting problem gambling among online sports and race bettors: Assessing the value of machine learning using behavioural and self-reported data.

Background and aims: Online gambling operators collect detailed behavioural data that can identify customers at risk of harmful gambling. However, there is limited clarity on how to optimally achieve this in practice, including which variables are most useful and whether short-term data windows are sufficient for risk detection. These details are increasingly important as regulatory frameworks emphasise timely intervention. We examined the value of machine learning in this context by comparing models trained on 30 days versus six months of behavioural data and exploring whether incorporating survey responses enhanced performance. Methods: Customers from two Australian sports and race betting sites (N = 1,470) completed a survey including the Problem Gambling Severity Index (PGSI) and measures of employment, income, gambling satisfaction, and number of gambling accounts. We built machine learning models to classify participants into risk groups (PGSI 1-7 [no-to-moderate-risk] vs. PGSI≥8 [high-risk]), comparing performance across data windows (30 days vs. six months), and with or without survey variables. Results: Models using only behavioural data achieved adequate classification accuracy (AUROC = 0.74-0.75), with similar performance across 30-day and six-month windows. The most predictive account-based variables were age, deposits per active day, average stake, and days since betting. Combining behavioural data with self-reported variables enhanced performance (AUROC = 0.76-0.85). Two self-reported variables-number of gambling accounts held and gambling satisfaction-were primarily responsible for these improvements. Conclusions: Machine learning models can detect at-risk customers on online sports and race betting sites using only 30 days of behavioural data. Performance can be improved by adding minimal, non-intrusive self-report measures.