Key Takeaways
- iGaming churn is cyclical, not binary. A 14-day gap means different things for different player types.
- Generic ML models trained on SaaS or subscription data systematically misfire on iGaming cohorts.
- AUROC, not accuracy, is the correct metric for churn prediction on imbalanced player populations.
- Lintvern achieved AUROC 0.85 on operator data, versus the 0.65-0.75 industry average.
In This Article
- 01The iGaming churn paradox
- 02What the data actually contains
- 03AUROC: the metric that matters
Most churn prediction models were built for subscription businesses: SaaS, streaming, telecoms. In those industries, churn is binary and lagged: a customer either renews or they don't, and you find out at renewal. iGaming is different. Players can go dormant for two weeks and reactivate at full value. Or they can deposit daily for a month and then never come back. The lifecycle patterns are fundamentally different, and a model that doesn't understand them will misfire constantly.
THE IGAMING CHURN PARADOX
In iGaming, a 14-day inactive period means very different things depending on the player. For a casual weekend bettor, it's normal. For a daily slots player, it's an early warning. For a high-roller who just lost a big session, it's a critical intervention window. Generic churn models treat inactivity as inactivity. Player-aware models treat it as context.
WHAT THE DATA ACTUALLY CONTAINS
Your warehouse holds the full behavioral signature of every player: session frequency, session duration, time-of-day patterns, game type distribution, stake variance, deposit frequency, withdrawal behavior, and bonus interaction history. A purpose-built iGaming churn model uses all of these signals, not just recency, frequency, and monetary value. The temporal graph structure of this data is particularly important: the sequence of events matters, not just the aggregate counts.
AUROC: THE METRIC THAT MATTERS
Most operators evaluate churn models on accuracy: what percentage of predictions were correct. This is the wrong metric for an imbalanced class problem where churners are 10-20% of the active base. AUROC (Area Under the ROC Curve) measures how well the model separates churners from non-churners across all possible thresholds. A random model scores 0.5. A good industry model scores 0.65-0.75. Lintvern's model scored AUROC 0.85 on our most recent operator data, meaning it correctly ranks a churner above a non-churner 85% of the time.
The gap between a generic churn model and a purpose-built iGaming model isn't marginal. It's the difference between sending a retention bonus to players who were going to reactivate anyway versus reaching the ones who are genuinely at risk. If your current model scores above 0.70 AUROC on your actual data, keep it. If it doesn't, you're burning retention budget on the wrong players.
Erik Ternav
Co-founder, Data & AI, Lintvern
Four years as a data scientist at IBM building AI solutions for enterprises across Europe. Now co-founder at Lintvern, where he solves the exact problem that killed most of those projects: getting models out of development and into the system where real decisions happen every day. In iGaming, that means daily player scores (bonus abuse risk, churn, promo sensitivity, VIP potential) delivered straight into the CRM operators already use.
More Articles
Bonus ROI
HOW BONUS ABUSE IS DRAINING YOUR GGR AND WHAT TO DO ABOUT IT
March 28, 2026
Architecture
THE CASE FOR WAREHOUSE-NATIVE SCORING IN IGAMING
February 28, 2026
CRM
CRM WITHOUT INTELLIGENCE: THE HIDDEN COST OF RUNNING BLIND
February 10, 2026
VIP
VIP UPLIFT MODELLING: IDENTIFYING HIGH-VALUE PLAYERS BEFORE THEY PEAK
January 22, 2026
FAQ
COMMON QUESTIONS
Questions about churn prevention scoring for iGaming operators.
What is player churn in iGaming?
Player churn in iGaming refers to active players becoming permanently inactive, ceasing to deposit, play, or engage with the platform. Unlike subscription businesses, iGaming churn is not defined by a cancellation event. It is typically defined operationally as a player who has not deposited or played within a rolling 30-, 60-, or 90-day window, depending on the operator's definition.
How is iGaming churn different from SaaS churn?
SaaS churn is binary and event-driven: a customer cancels their subscription on a specific date. iGaming churn is probabilistic and behavioral: a player gradually disengages through declining session frequency, smaller deposits, and narrowing game selection. Players can also appear to churn and then reactivate, making the label inherently ambiguous. Purpose-built models treat churn as a continuous risk score rather than a binary state.
What is AUROC and why does it matter for churn prediction?
AUROC (Area Under the ROC Curve) measures how well a model separates two classes (churners and non-churners) across all possible decision thresholds. A random model scores 0.50. A good industry model scores 0.65-0.75. Lintvern's model scored 0.85 on recent operator data. Accuracy is a misleading metric for churn because it rewards models that predict 'no churn' for everyone in a population that is 85% non-churners.
How far in advance can a churn model predict a player leaving?
A well-built iGaming churn model can identify at-risk players 14-30 days before they become fully inactive, based on leading behavioral indicators like declining session frequency, narrowing game selection, and reduced deposit amounts. This intervention window is enough to trigger a targeted retention campaign before the player disengages completely.
What data does a churn prediction model need?
The minimum viable dataset includes player sessions (timestamps, duration, game type), deposit history, withdrawal history, and bonus interactions, covering at least 6 months. Richer models also incorporate bet-level transaction data, time-of-day patterns, and device or channel data. All of this typically exists in the operator's data warehouse.
READY TO SCORE AND
PROTECT YOUR REVENUE?
Partner with Lintvern to cut bonus waste, predict churn, and make your existing CRM smarter, without changing a single tool.