Customers only: preview

VIP cohort drift: how to validate a churn model before trusting its scores

Evaluate iGaming churn models through label definitions, temporal validation, calibration and cohort monitoring without confusing prediction with causation.

For CDOs, Heads of BI and CRM analytics directors

A churn probability is only meaningful relative to a defined outcome, an observation window and a population. Without those three elements, a score can look sophisticated while remaining impossible to evaluate.

Executive summary

Define what churn means before evaluating a model. A thirty-day absence of settled wagers, no authenticated sessions and a closed account are different outcomes. Combining them into a single label can obscure both commercial interpretation and service responsibilities.

The framework here concerns analytical validation and account-service understanding. It does not use losses, possible vulnerability or predicted inactivity to prescribe personalised gambling inducements. The model must support a reviewable observation, not create a false certainty about a person’s future behaviour.

1. Specify the population, label and time boundaries

Freeze the cohort before observing its outcome

A VIP cohort needs a membership rule and an effective date. If membership is assigned using future revenue, the evaluation already contains information unavailable when a prediction would have been made. Preserve historical membership instead of overwriting it with the latest tier.

  • Define the observation window and prediction timestamp.
  • Define the outcome and the period in which it is measured.
  • Exclude information created after the prediction timestamp.
  • Distinguish restricted, self-excluded, closed and technically inaccessible accounts.
  • Identify cases without enough follow-up to determine the outcome.

Avoid turning missing data into inactivity

A delayed gameplay feed can make an active population appear absent. Account migration can change identifiers. A payment outage can interrupt activity without indicating a change in preference. Feature freshness and identity continuity therefore belong in the evaluation dataset.

Keep successful deposits separate from failed attempts, and record whether sessions or wagers are the relevant exposure. RTP variance can affect short-window financial observations; it should not be treated as evidence of future behaviour on its own. At this stage, the purpose is to establish what was observed reliably, not select a promotional response.

Customers only

The rest of this guide is for Bounty AI customers.

Full guides, product screens and walkthroughs are available to Bounty AI customers and pilot partners. Access is by invitation, and customer sign-in is not open yet.

  1. 2. Evaluate out of time and distinguish ranking from calibration
  2. 3. Monitor drift without confusing it with model failure
  3. Operational outcome: fewer decisions based on misleading certainty
Pilot on your data

Pilot partners get the full library when their pilot starts.

More in the Client library

All guides