The models the CRM team runs
The oldest and most profitable use of machine learning in online gambling is deciding what to do with each customer. An operator with a million accounts cannot treat them individually by hand; it can score them, group them, and let the scores decide which message, offer, game or intervention each receives. This lesson covers the customer-value models (churn, lifetime value, propensity), the personalisation models (recommendation, next-best-action) and the experimentation discipline that keeps them honest, and it closes with the constraints that regulation puts on all of them.
Churn
A churn model predicts, for each active customer, the probability that they will stop playing within a horizon (thirty, sixty or ninety days). It is trained on the histories of customers who did and did not lapse, with features drawn from recency (days since last bet, last deposit, last login), frequency and monetary value, trend (is activity falling), product mix, promotional response, support contacts and session patterns. Its output ranks customers by risk of leaving, and the CRM team acts on the ranking: a retention offer to the high-risk high-value segment, a content message to the high-risk low-value one, nothing to the customers who will stay anyway.
The value of a churn model is in the targeting. Blanket retention offers cost money on customers who did not need them; a model that identifies the ones who do, and the ones who can be saved, cuts the cost per retained customer sharply. The evaluation is an experiment: hold out a random group from the model-driven treatment and measure the difference in retention and value, which is the only way to know that the model changed behaviour rather than predicted it.
Lifetime value
A lifetime value model estimates the net revenue a customer will produce over a future period, usually from early signals: first-week deposits, product, acquisition channel, device, market, early session behaviour. It is used for acquisition (bid more for channels and campaigns that bring high-value customers; set affiliate CPA by predicted value), for early-life CRM (invest in customers likely to become valuable before they are), and for the VIP pipeline (identify future high-value customers early). The model's error is large for any individual and useful in aggregate, and its main risk is the feedback loop: a model that predicts value from early deposits, used to decide who receives offers, produces customers who deposit more because they received offers, which the next model learns as value.
Propensity and next best action
Propensity models score the probability of a specific behaviour: depositing this week, trying a new product, responding to a particular offer, reactivating after a lapse. A next-best-action system combines propensities, value and business rules to choose, for each customer at each moment, which of the available actions (an offer, a message, a recommendation, a pause, nothing) has the highest expected value net of cost and within the constraints. It is the engine behind personalised CRM at scale, and its design is where commercial optimisation meets player protection: the same system that chooses the best offer must know which customers may not receive one.
Recommendation
Casino lobbies with thousands of games and sportsbooks with thousands of markets need ordering, and recommendation models supply it: collaborative filtering (customers who played this also played that), content-based ranking (games with similar features to what this customer plays), and bandit algorithms that balance showing what is known to work with testing what might. Recommendation moves engagement measurably, particularly for new customers who do not know the catalogue, and its constraints are regulatory as much as technical: several markets restrict how games and offers may be presented to customers showing signs of harm, and a recommender that pushes higher-volatility games to a customer who is chasing losses is a compliance failure in the making.
Personalised promotions
The bonus engine can award anything; the models decide what. Personalised promotions set the offer type, amount, wagering and timing per customer from the propensity and value models, subject to the market's rules and the operator's margin targets, and they are the most scrutinised application of machine learning in the sector. Regulators in Britain, the Netherlands and Australia have questioned whether personalised offers exploit vulnerability, several enforcement cases have found offers made to customers who should have been protected, and the rules on inducements in Ontario and elsewhere limit what can be personalised at all. The operating principle that survives regulatory review is that protection rules are applied before commercial models, not after: a customer flagged for harm, on a time-out, self-excluded or under an affordability restriction is removed from the promotional population before the next-best-action engine sees them.
Experimentation
None of the above can be trusted without experiments. A model that predicts churn does not prove that acting on the prediction retains customers; a recommender that raises clicks may not raise value; a personalised offer may cannibalise deposits that would have come anyway. Operators run controlled experiments (randomised holdouts, A/B tests, uplift modelling that predicts the effect of treatment rather than the outcome) and measure against the metrics that matter (net revenue, retention, cost per retained customer) over long enough windows to see the tail. The experimentation platform, the discipline of running enough tests with enough power, and the culture of accepting null results are what separate operators whose models work from those whose models merely exist.
Data and features
All of these models run on the same foundation: an event stream with consistent identifiers, a feature store of customer attributes computed on a schedule and in real time, labels (who churned, who responded, who was harmed) recorded accurately, and a training pipeline that retrains as behaviour drifts. The platform course covers the architecture; the point here is that the models are only as good as the labels, and that a churn definition, a value definition or a harm label that changes silently invalidates every model built on it.
The constraints
Customer models in gambling operate under three constraints that do not apply in retail. Data protection: profiling requires a lawful basis, transparency and, for decisions with significant effects, human oversight and the right to contest, under the GDPR and its equivalents. Player protection: commercial models must be subordinate to harm models, with suppression lists applied first and the interaction between them audited. Marketing rules: personalised offers are still inducements, subject to consent, opt-in by product and channel where required, advertising codes and inducement bans. A CRM model that ignores any of the three is an enforcement case.
What to take from this lesson
Churn, lifetime value and propensity models rank customers for CRM action; next-best-action systems choose among actions; recommenders order catalogues; personalised promotions set offers per customer. Their value is proven only by controlled experiments against net revenue and retention. They depend on consistent event data, feature stores and accurate labels. And in gambling they are constrained by data-protection rules on profiling, by the primacy of player-protection models, and by marketing and inducement rules, with protection applied before commercial optimisation.