Models that protect the margin
Every online gambling operator is attacked every day: by bonus-abuse rings, by stolen-card fraud, by account takeover, by money launderers, by match-fixers and by professional bettors who are not fraudulent but are unprofitable. The rules-based controls that catch the obvious cases (velocity checks, closed-loop payments, limit enforcement) are the first line; machine learning is the second, and it is where the sophisticated cases are found. This lesson covers the risk models by family, how they are built and evaluated, and the escalating contest with adversaries who now use the same tools.
Fraud
Bonus abuse and multi-accounting. The commonest fraud in casino: one person, many accounts, each claiming a welcome bonus and playing it down to the cashout cap with low-risk strategies. Models link accounts through device fingerprints, IP and network signals, payment instruments, behavioural patterns (the same betting sequence on the same game at the same stake), registration data similarities and timing. Graph-based approaches, which treat accounts, devices and payment methods as nodes and their shared attributes as edges, find rings that pairwise rules miss, and are the standard for this problem.
Payment fraud. Stolen cards, friendly fraud (chargebacks by the genuine cardholder), and synthetic identities. Models score transactions on card and customer history, velocity, device, geolocation consistency, amount patterns and issuer signals, alongside the payment provider's own scoring, and route high-risk transactions to step-up authentication or review. The evaluation metric is the trade-off between fraud losses and false declines, each of which is measured in revenue.
Account takeover. Credential stuffing, phishing and session hijacking against existing accounts, followed by a withdrawal to a new method. Models detect the takeover sequence (new device, password change, new withdrawal method, withdrawal request, in minutes) and behavioural anomalies (a customer who bets differently), and trigger re-authentication or holds. Behavioural biometrics, which model how a customer types and moves, are the newer signal.
Arbitrage and bonus-hunting in sports. Customers who use promotions and price errors systematically. Models identify the patterns (bets on both sides across operators, stakes that exactly clear a free bet, activity concentrated on boosts) and feed the stake-factoring decision covered below.
Anti-money-laundering
Anti-money-laundering rules require operators to monitor transactions and behaviour for suspicion, and transaction monitoring is now model-driven at scale. Rules catch thresholds (deposits above an amount, structuring below it); models score the pattern: deposits followed by minimal play and withdrawal, mismatches between declared occupation and spend, use of multiple payment methods, links to accounts under investigation, activity consistent with typologies the financial intelligence unit has published. Alerts are prioritised for the compliance team, whose decisions (enhanced due diligence, source-of-funds enquiry, suspicious activity report, account closure) become the labels the next model learns from. Regulators expect the monitoring to be risk-based, documented, tuned and explainable, and enforcement cases have found operators whose systems generated alerts nobody reviewed, which is a model-governance failure rather than a model failure.
Sports integrity
Sportsbooks see betting patterns that nobody else does, and the integrity function models them: unusual volumes on a market relative to its liquidity, price movements against the market, bets from accounts with no history in a sport, concentrated action on lay outcomes or on obscure events, and correlated activity across accounts. Alerts are shared with integrity bodies and sports governing bodies under the information-sharing arrangements that regulated markets require, and they have been the evidence in match-fixing prosecutions. The models are anomaly detectors on a base of market-level statistics, and their value depends on the sportsbook's coverage: the more markets and customers, the better the baseline.
Sharp detection and stake factoring
Professional and skilled bettors are not fraudsters, but their bets lose operators money, and every sportsbook models them. Closing line value (how often a customer's price beats the closing price) is the classic measure; models add bet timing (early money on soft lines), market selection (obscure leagues, player props), stake patterns (maximum stakes at specific prices), correlation with known sharp accounts and, increasingly, cross-operator signals. The output feeds stake factoring: limits on what a customer may bet, applied automatically by the bet-acceptance engine. The practice is lawful and universal, and it is the most contested customer-facing use of models in sports betting, with regulators in several markets examining whether restrictions are applied fairly, whether customers are told, and whether the same operators that restrict winners are equally quick to restrict losing customers showing harm.
Building and evaluating risk models
Risk models share a set of engineering problems. Labels are scarce and delayed: fraud is confirmed weeks after the transaction, laundering may never be confirmed, and the labelled cases are the ones the old system caught, which biases the new one. Class imbalance is extreme: a fraction of a per cent of transactions are fraud, so accuracy is meaningless and the metrics are precision and recall at the operating point, with the cost of each error type set by the business. Adversaries adapt: a model that works changes the behaviour it detects, and drift monitoring and retraining are permanent. Explanations are required: a compliance analyst must be able to see why an alert fired, and a customer whose account is restricted has, under data-protection law, a right to meaningful information about the logic.
The operational design that works is a layered one: rules for the certain cases, models for scoring and prioritisation, human review for decisions with significant effects, and a feedback loop from review outcomes into training data. Fully automated adverse decisions (closing an account, confiscating funds) are avoided both for accuracy and because data-protection and gambling rules require human involvement.
The adversary has models too
The 2026 risk assessments published by regulators name the new threat directly: AI-generated identity documents and deepfake video defeating verification, synthetic identities built from generated data, automated bot play, and generated content in affiliate and phishing campaigns. Identity vendors respond with liveness detection, document forensics trained on generated fakes, and behavioural signals that generated documents cannot supply; operators respond with re-verification triggers and layered checks. The contest is now between models on both sides, and the operator's advantage is its own data: the history of its genuine customers is the baseline that synthetic ones fail to match.
What to take from this lesson
Risk models cover bonus abuse and multi-accounting (graph methods on shared attributes), payment fraud (transaction scoring against false-decline cost), account takeover (sequence and behavioural anomaly detection), anti-money-laundering (typology scoring with human review and a feedback loop), sports integrity (market anomaly detection shared with integrity bodies) and sharp detection (closing line value and pattern signals feeding stake factoring). They face scarce delayed labels, extreme imbalance, adapting adversaries and explanation duties, and are built as layers of rules, models and human review. Adversaries now use generative tools against identity checks, and the operator's own customer data is the defence.