New paper out today in IGS, using machine learning on online gambling data from @BCLC ’s @PlayNowCasino to predict self-exclusion as a marker of gambling problems. Finkenwirth, MacDonald @tilmanlesch @LukeClark01 @UBCPsych @DMCbrainhealth. Thread 1/n https://twitter.com/ubcnews/status/1325838119363555328

Paper here https://www.tandfonline.com/doi/full/10.1080/14459795.2020.1832132 Shareable here https://www.tandfonline.com/eprint/5STBHNG4ZYVDPHHQBSKR/full?target=10.1080/14459795.2020.1832132
& preprint on our website https://cgr.psych.ubc.ca/publications/ 

This is one of 1st studies looking at behavioural tracking of online gambling in North American data. Although online gambling has high potential for monitoring & detecting risk, key challenge for building predictive algorithms is knowing which gamblers actually have problems

We used 1 y de-identified data from eCasino (30k gamblers, half a billion bets = big). Targets: 2,157 w record of self-exclusion, vs 17,526 controls. 20 behavioural inputs eg days gambled, session length, # games. Machine learning was random forest with 10 fold cross-validation

2 key findings: 1) primary model had 75% accuracy (AUROC) in classifying self-excluders. Follow-up models tested some key decisions (balancing, data inclusion threshold): AUROCs 65-76%. We ran a logistic regression ‘benchmark’ (only 39% - but better in the follow-up analyses)

2) which #gambling variables were most predictive? In a smaller set of uncorrelated inputs, variance in amount bet per session accounted for 32% of signal. We think v interesting: variable more than high spend linked to gambling problems. Sign of loss chasing?

We see these findings as promising 'proof of principle' for detecting high-risk gamblers online. ‘Bet by bet’ data is just one of several kinds of data that online operators have access to (e.g. account deposits, demographics, using RG resources...) to further improve performance

But many limitations and future questions. Our model is predicting self-exclusion; in other work ~3/4 of self-excluders have gambling problems, but it is not a perfect ‘marker’ for gambling problems. Plus, we are using ‘ever’ record of VSE not ‘prospective’ self exclusion

Huge effort by CGR data science peeps, & our study builds on much important prior work in this area including @howard_shaffer @debilaplante @heathermarygray @kahlilphilander @simodragicevic @michaelJAwohl @LuquiensA @DrSalGainsbury.