A propensity model is a predictive model that estimates the probability that a specific customer will take a particular action within a defined window — deposit again this week, churn this month, respond to an offer, become high value, or convert after a free trial.
Definition
A propensity model is a predictive model that estimates the probability that a specific customer will take a particular action within a defined window — deposit again this week, churn this month, respond to an offer, become high value, or convert after a free trial. It is trained on historical data where the outcome is known, learning which combinations of behavioural, demographic and contextual features precede the action, then scores current customers so activity can be targeted at those most likely to respond.
A propensity score is a probability, not a certainty, and it reflects patterns in past data, so it can be wrong for individuals and can carry forward biases in the training data. It also predicts who is likely to do something, not whether an intervention changes that — a customer with a high deposit-propensity score may deposit anyway, so pairing propensity with incrementality thinking avoids spending on people who need no nudge.
In context
For iGaming operators, propensity models drive several CRM decisions: churn-propensity scores prioritise retention outreach, deposit-propensity scores time reload offers, value-propensity scores flag new players likely to become significant so onboarding can be tailored, and offer-response scores decide who receives a given promotion. Used well, they make marketing more efficient by concentrating effort where it is likely to matter.
The responsible-gambling constraints are strict and specific to this sector. The same features that predict a valuable, engaged player — rising deposit frequency, longer sessions, chasing behaviour after losses — also predict gambling harm, so a high deposit- or value-propensity score can be flagging a player at risk.
Propensity outputs must be gated by harm and affordability checks: a player the model rates as highly likely to deposit again, who also shows risk indicators, should be routed to a safer-gambling intervention, not a marketing push. Self-excluded and at-risk players must be excluded from all propensity-driven targeting.
Regulators have criticised operators for using predictive models to identify and market to players whose behaviour indicated harm. Done responsibly, propensity modelling improves relevance and efficiency; done without the safeguards, it optimises toward extracting more from the players least able to sustain it, which is both an ethical failure and a regulatory one.
For affiliates, propensity work is almost entirely operator-side, but it is part of why operators value predictable, sustainable player cohorts over volatile high-spend ones.
Worked example
An operator's churn-propensity model flags 4,000 players as high-risk-of-lapsing this month. Before any are messaged, the list is filtered against responsible-gambling and affordability flags; 300 are removed and routed to welfare checks.
The remaining players receive a light re-engagement message, and an incrementality holdout confirms the campaign's true effect.
Related terms
Frequently asked questions
Browse the full iGaming & affiliate glossary — hundreds of EN/RU terms with examples.
← Back to glossary