Sam Andersson

A machine-learning study of 4.5 years of individual online-casino account data, examining how early and how reliably problem gambling can be flagged — using leakage-controlled, temporally held-out validation and 30/60/90-day truncation analysis.

Develops a held-out total-variation certificate for finite-basis drifting models by expressing finitely sampled numerator drift as vec(V_X) = M c. The certificate propagates field-estimation error, operator-calibration error, representation residuals, and probe conditioning, and abstains when the observability margin is nonpositive. The analysis also identifies rank and symmetry degeneracies and proves large-bandwidth collapse toward mean matching.

Combines heavy-tail-aware window representations, teacher–student CVAE embeddings, backlog-aware label inference, and Gaussian hidden Markov models to derive a three-level operational risk proxy from routine online gambling telemetry under sparse, selectively observed manual labels. Evaluation uses leakage-controlled holdouts and capacity-constrained alerting across transactions, bets, sessions, and payments.

Pairs a calendar-time representation of 14.2 million player-days with a Bayesian mixed-emission hidden semi-Markov model of active gambling bouts. The model separates latent regimes that can share similar turnover while differing in playing time, approved deposits, and session counts, and uses Sweden's temporary deposit-cap period as a descriptive application rather than a causal policy-effect estimate.

Applies machine-learning methods to longitudinal clinical and epidemiological data to predict which women go on to develop depressive symptoms postpartum.