Machine Learning Flags Gambling Risk From Routine Betting Data Without Clinical Labels

Researchers have built an end-to-end pipeline that turns routine online gambling telemetry into an auditable, three-level risk flag, and they have reported exactly how well it works, including where it fails. The study, published in EPJ Data Science, analyzed de-identified records from a single licensed online gambling operator across four behavioral streams: transactions, bets, sessions, and payments. The goal was not a clinical diagnosis but an operational proxy definition of gambling-related risk, designed for triage and workload planning under finite review capacity.

The motivation is structural. Gambling telemetry is heavy-tailed, meaning a small number of extreme days account for a large share of wagering or losses, and rule-based responsible gambling triggers, such as fixed deposit thresholds, can flag risk only after substantial losses have already accumulated. Manual analyst assessments, the closest thing to a reference standard, are sparse and selectively observed: the probability that a customer gets reviewed depends on the very behavioral signals the system is trying to measure, a missing-not-at-random problem compounded by annotation backlogs and feedback loops from existing responsible gaming systems.

The pipeline addresses these issues in stages. Behavioral time series were converted into rolling 30-day windows with a one-day stride, and a strict leakage audit excluded any window overlapping its label date, enforcing causal ordering. Extreme behavior was represented with exceedance indicators, whether unusually large events occurred, and exceedance magnitudes, how far observations exceeded training-derived thresholds, summarized over short horizons so that frequency and severity of extremes stay separate rather than collapsing into scale-normalized averages. Additional features captured sustained momentum versus short-lived spikes and short-horizon cascades across behavioral dimensions.

Representation learning used a hierarchical conditional variational autoencoder with three parallel temporal encoders covering short, mid, and long receptive fields, trained in a teacher-student regime. The teacher learned from abundant responsible-gaming proxy signals; the student was then fine-tuned on sparse manual analyst labels with its encoders frozen, adapting only the decision boundaries. To correct selective labelling, the authors introduced backlog-aware label inference, a hazard model of labelling propensity combined with a Bayes adjustment that pseudo-labels high-confidence unlabeled training windows. Pseudo-labels were confined to the training split; all evaluation used original manual labels only. Finally, a regularized Gaussian hidden Markov model treated the embeddings as noisy observations of a slowly evolving latent regime, producing calibrated three-class probabilities.

The headline agreement results are moderate and stream-dependent. On labelled test windows, balanced accuracy ranged from 0.380 for transactions to 0.624 for payments, against a three-class chance level of roughly 0.333. Bets achieved the strongest macro-averaged F1 at 0.543, and sessions had the highest high-risk recall at 0.757. The payments figure demands caution: test support was heavily skewed, with 514 low, only 22 medium, and 1386 high-risk windows, and high-risk recall was just 0.025, so the balanced accuracy was driven almost entirely by low- and medium-class performance. Transactions, the weakest stream, had only 3.0% original training-label coverage, and a fixed-representation diagnostic showed balanced accuracy rising as more labelled groups were used, indicating the near-chance result is primarily label-limited rather than evidence that transaction data are uninformative.

The most operationally relevant comparison was a capacity-matched test. On a restricted common-support customer-week set where the model, the operator’s existing responsible-gaming score, and a random queue could all be applied under the same top-10-per-week budget, the model detected more subsequently observed escalations in every stream: 65% versus 52% and 42% for transactions, 55% versus 43% and 31% for sessions, 70% versus 61% and 46% for payments, and 78% versus 68% and 54% for bets. Median alert-to-observed-assessment intervals ranged from 65 to 210 days, but the authors are explicit that these are not times to harm. They reflect alert timing relative to selectively observed manual assessments under a backlogged review process, and a customer can remain highly ranked for a long period before review.

Yield under realistic budgets was sobering. At a 30-day horizon, alert-level positive predictive value was low, between 0.8% and 4.9%, and escalation-level recall ranged from 21% to 43%, reflecting low base rates and finite review capacity. Calibration varied by stream; temperature scaling reduced expected calibration error but residual miscalibration remained, motivating periodic recalibration and drift monitoring. In a frozen-pipeline same-operator temporal holdout, balanced accuracy differed from the combined labelled evaluation set by only minus 0.02 to plus 0.02 across streams, suggesting temporal portability within this operator, though this is explicitly not validation across independent operators, jurisdictions, or regulatory contexts.

The authors also report honest negative and heterogeneous findings. Label augmentation effects differed by stream, improving balanced accuracy in transactions, sessions, and payments but not bets. Against conventional non-temporal baselines, the full pipeline was strongest in bets and competitive in payments, while some baselines matched or exceeded it in sessions and transactions. An incumbent responsible-gaming proxy alone achieved comparable or higher agreement in transactions and payments, which the authors interpret as possible coupling between operational labelling and existing responsible gaming systems. Threshold sensitivity analyses over acceptance thresholds from 0.50 to 0.90 showed the qualitative conclusions did not depend on a uniquely favorable setting, though transaction regimes were comparatively unstable.

The limitations are stated plainly. The data come from a single operator, so generalization requires validation elsewhere. Manual analyst assessments are an operational reference standard, not clinical diagnoses, and may be influenced by the systems under study. The timing analyses are offline simulations anchored to observed assessments, not estimates of intervention effectiveness or causal early warning, and the direction of triage-coupled observation bias is ambiguous. The multi-stage pipeline also complicates attribution of performance to any single component. Future work the authors identify includes multi-operator validation, prospective evaluation of intervention policies, linkage to external outcomes such as treatment contact, and formal audits for fairness, drift, and feedback effects.

The broader significance lies in the framing. Rather than treating risk as a static classification target or chasing a single headline metric, the framework treats it as a latent, time-varying regime manifested through clustered extremes, enforces causal ordering at the window level, corrects supervision under selective labelling, and translates probabilities into explicit, auditable decision policies such as thresholds, top-K queues, and persistence gates. The authors argue this problem structure, heavy-tailed digital traces, evolving latent states, selectively observed human labels, and capacity-constrained decisions, extends well beyond gambling to fraud monitoring, insurance claims triage, credit risk, and platform integrity. Within gambling monitoring, the contribution is a transparent definition framework whose utility is judged by capacity-constrained operating points, with class imbalance and selective-observation limits made explicit rather than hidden behind aggregate accuracy numbers.

Subject of Research: Deriving an operational proxy definition of problem gambling risk from routine online gambling data using heavy-tail-aware representation learning and dynamic Bayesian state modelling

Article Title: Heavy-tail-aware representation learning and dynamic Bayesian state modelling to derive an operational proxy definition of problem gambling risk from routine online gambling data

Article References: Andersson, S., Westerlind, H., Koski, T., Lyon, K., Carlbring, P., Lindner, P., & Molander, O. (2026). Heavy-tail-aware representation learning and dynamic Bayesian state modelling to derive an operational proxy definition of problem gambling risk from routine online gambling data. EPJ Data Science, 15(1), Article 77. https://doi.org/10.1140/epjds/s13688-026-00698-3

Image Credits: AI Generated

DOI: 10.1140/epjds/s13688-026-00698-3

Keywords: problem gambling, online gambling, machine learning, hidden Markov model, variational autoencoder, heavy-tailed data, responsible gambling, missing not at random, behavioral telemetry, risk monitoring, capacity-constrained alerting, EPJ Data Science