Player Churn Rate in iGaming: Measure, Predict, Reduce

Define churn, calculate the rate, spot early signals, predict who leaves and prove with a control group that your retention work changes the outcome.

Updated 8 November 2026 · 11 min read · Gamble Grip

Player churn rate is the share of players who were active in one period and are not active in the next. The formula is simple. The hard parts are what counts as active, which window you use, and whether your retention work changes anything that would not have happened anyway.

This guide takes you through five steps: define churn, calculate it, read the early signals, predict who will leave, and prove with a control group that your intervention worked. It is written for CRM and retention leads at casinos and sportsbooks.

Why churn is an inference

In subscription software the customer cancels. A casino or sportsbook has no cancellation event. The player stops depositing and betting, and the account stays open. You infer churn from inactivity, so you must define it yourself.

That definition drives every number downstream. Two teams can report 25% and 45% churn for the same player base and both be right, because one counts a login as activity and the other requires a wager.

Define churn before you count

Decide four things once and write them down.

Activity event. A wager or deposit is more honest than a login, because players log in to check a balance or withdraw.
Scope. The whole account or one product. A player who left slots for sports has left casino but not the brand.
Window. 30, 60 or 90 days of inactivity.
Population. Players with at least one deposit. Registrations without a deposit belong to onboarding and first-deposit analysis.

Merchie and Ernst, in a study on data from the Belgian online gambling company Gaming1, used a concrete version. A player is a churner after 35 days without activity. A day counts as active if the player had at least one gaming session, sports bet or deposit. Copy the shape, not the values: an explicit activity event and an explicit threshold.

Keep two states apart. A lapse is a short silence that triggers a message. A loss is a long silence, for example 90 days, and is the number you report. A player past your inactivity threshold is dormant. Bringing dormant players back is a separate task, covered in player reactivation campaigns.

Calculate churn rate and value churn

The calculation uses a base period and an observation period.

Churn rate = players active in the base period who are inactive in the observation period ÷ players active in the base period
Retention rate = 1 − churn rate
Value churn = NGR in the base period from players who churned ÷ total NGR in the base period

Example. An operator counts 20,000 depositing players with at least one wager in January. In February 13,000 of the same players bet at least once and 7,000 do not. Churn rate is 7,000 ÷ 20,000 = 35%. Retention rate is 65%.

Three rules keep the number honest.

Use the same players in the numerator and the denominator. New February registrations do not belong in the calculation.
Report a short-window lapse rate and a long-window loss rate side by side. Some of the 7,000 will return in March.
Cut the rate by cohort: registration month, acquisition channel, first-deposit size and first product. The method is in cohort analysis in iGaming.

Churn rate and retention rate are the same measurement from two sides. See the glossary entry on churn rate.

Weight churn by value

A flat rate treats a player who deposits €10 a month and a high-value regular as equal. Value churn corrects this. If 35% of players churned but they produced only 12% of NGR, the problem is smaller than it looks. If they produced 30%, you are losing core revenue.

Same head-count churn, different damage

Example. Share of base-period NGR lost with churned players.

Use NGR rather than GGR, since bonus cost can turn a busy player into a loss. The difference is explained in NGR vs GGR.

Read a high rate against your own history

A single churn figure is comparable only when the activity event, window and population match. Use these checks instead of an outside benchmark.

Your own trend with the definition frozen. A rise of several points over two or three comparable periods is a signal.
Cohort curves. If recent cohorts lose players faster than older ones did at the same age, acquisition quality, onboarding, product or bonus policy changed.
Value churn against head-count churn.
Seasonality. Major sporting events, payday cycles and holidays move activity. Compare like with like.

Find the early signals

Players rarely vanish overnight. Behaviour decays first, and a retention programme can act on that decay. Turn these signals into numbers and test them on your data.

Signal	What to measure	Why it matters
Longer gaps between sessions	Current gap against the player's own median gap	A personal baseline beats a fixed threshold
Falling deposit frequency	Deposits in the last 14 days against the previous 14	Money in is the earliest commercial signal
Shrinking stakes or sessions	Rolling average against the 60-day average	Interest fades before activity stops
Narrowing product mix	Products used per month	Fewer products, fewer reasons to return
Withdrawal with no redeposit	Time from withdrawal to next deposit	The player cashed out and left
Friction events	Failed deposits, KYC delays, support tickets	Often a fixable cause
Ignored messages	Open or click rate per player	Channel fatigue
Bonus dependence	Share of activity inside bonus periods	Engagement may not survive the offer

Swipe the table sideways →

The signal that matters is change against that player's own history. A weekly player and a daily player have different normal gaps.

Causes differ by operator, so treat these as hypotheses to test. Common ones: no second deposit, bonus-led acquisition, dependence on one league or tournament, friction in payments and verification, irrelevant messages, and competitor offers. To find which dominate in your base, map where players drop out by lifecycle stage, as in player lifecycle management.

Predict who will leave

A rule such as "30 days without a bet" acts only after the player is gone. A model scores players while they are still active, so a fix or a message can arrive earlier and a limited budget can go to the right players.

Define the target

The target has three parameters. Getting them wrong is the most common modelling mistake.

History window: how far back the model looks, for example 60 days.
Churn definition: inactivity of N days.
Prediction horizon: how far ahead you predict. In the Merchie and Ernst study the model read 60 days of history and predicted churn over the next 30 days.

Score only players who are still active on the scoring date. Predicting that a dormant player is dormant is trivial.

Choose features and a model

Start from the signal table. Recency (days since last wager and last deposit), frequency (sessions and deposits per week), value (stakes, deposits, NGR), trends (last 14 days against the previous 14), product mix, payment events, bonus share, support contacts, message engagement and tenure.

Begin with a regularised logistic regression. It is fast, a CRM manager can read it, and it sets a baseline that other models must beat. Then try gradient-boosted trees. Sequence models learn directly from daily time series.

Merchie and Ernst used recurrent networks on the Gaming1 data. Their results set realistic expectations.

About 22% of the test samples were churners.
Three recurrent cell types reached precision of 0.62 to 0.64 and recall of 0.49 to 0.61 for the churn class. Accuracy was about 0.83.
Accuracy alone misleads on imbalanced classes. A model that says "no churn" for everyone would score about 78% on that split.

The model ranks players well enough to be useful. It is also wrong for a large minority of them.

Validate like a forecast

Split by time. Train on earlier dates and test on later ones. A random split leaks future behaviour into training.
Check for leakage. Any feature that uses information after the scoring date makes the model look strong and fail live.
Use ranking metrics. Measure precision and recall at the cut-off you will use, and top-decile lift: how much higher the churn rate is in the top 10% of scores than on average.
Check calibration. If the model says 70%, about 70% of those players should churn.
Plan for drift. In the same study, the authors found that recall of churn events starts to drop when the training window is shorter than about one year. They also saw the metrics shift around two one-off events: the operator stopped several bonus types for legal reasons in March 2020, and the casino was unreachable for several days in June 2020. Re-check monthly and after any major product or bonus change.

A manager reads a gauge that ranks figures on a belt, while a separate lever sits unpulled beside it — A churn score ranks players. It does not say who will respond to a message.

A model tells you who is likely to leave. It does not tell you who will respond to a message. Some high-risk players are lost anyway, some would stay without contact and a few are irritated by it. Uplift modelling addresses this. Gutierrez and Gérardy define it as the set of techniques used to model the incremental impact of an action or treatment on a customer outcome. You can start without uplift models, but you need a control group.

Intervene and prove the effect

Match the action to the cause

Fix friction first. Failed deposits, slow withdrawals and verification delays are cheaper to fix than to compensate. Route these players to support, not to a promotion.
Trigger on behaviour. Contact a player when the gap exceeds that player's own norm, not on day 30 for everyone. The wider playbook is in player retention strategies.
Tie offers to the product the player used. Test a product-specific offer against a generic reload.
Rotate channels on evidence. If a player ignores one channel, a fourth email is not the answer.
Limit bonus dependence. Heavy discounting lifts activity this month and trains players to wait for the next offer. Compare NGR after bonus cost.
Give valuable players a personal contact from a VIP or retention manager. Automate the long tail.

Run the holdout test

Reporting that 30% of messaged players returned proves nothing, because some of them would have returned anyway.

Define the eligible audience before the campaign, for example players above a risk cut-off who are active and not suppressed.
Assign a random holdout from that audience. The smaller the base, the larger the share you need.
Treat only the rest. The holdout receives no message from this programme.
Compare return rate, deposits per player and NGR per player after bonus and campaign cost over a fixed window.
Report the difference, with a confidence interval if volumes allow.

Example. An operator selects 10,000 eligible at-risk players. 1,000 go to the control group and 9,000 are treated. In 30 days 2,430 treated players deposit (27%) and 190 control players deposit (19%). NGR per player is €14.20 treated and €11.00 control. Cost per treated player is €1.50.

The depositor rate rose by 8 percentage points. NGR per player rose by €3.20, or €1.70 after cost. Across 9,000 players that is €15,300. A raw "27% returned" overstates the effect, because 19% returned with no contact.

Share of at-risk players who deposited within 30 days

Example. 9,000 treated players, 1,000 control players.

Four things spoil the test.

A control group chosen by hand or different in value from the treated group. Randomise, then check past NGR and tenure.
A holdout contaminated by other campaigns. Keep a persistent global holdout when several programmes run at once.
A window that is too short. A bonus can pull a deposit forward without adding one. Look at 30 to 60 days.
Different player IDs in the CRM and the PAM. Audience, IDs and NGR must reconcile before launch.

If a true holdout is impossible, a matched historical cohort is a weaker substitute and must be labelled an estimate. You can model cohort economics in advance with the campaign economics calculator.

Churn work and responsible gambling

Churn signals and harm signals overlap. A churn programme must not chase vulnerable players.

Stop direct and targeted marketing and the take-up of new bonus offers where strong indicators of harm are identified. Limit direct marketing further as indicators strengthen. Indicators include chasing losses, erratic betting, use of time-outs, previous self-exclusions and failed deposits. Several of these look like ordinary at-risk behaviour on a churn dashboard. A model trained only on revenue can rank such players as top priority for a bonus.

Situations outside the account, such as financial difficulty, divorce, bereavement and job loss, also matter. Act on the information you hold.

Rules for the operator:

Apply suppression before scoring or targeting: self-excluded, time-out, cooling-off, harm-flagged and unsubscribed players never enter the audience.
Respect consent per channel.
Keep the safer gambling team in the loop on audience definitions and log who was contacted and why.
Never optimise a model for "most likely to keep playing" without a harm filter.

See responsible gambling in CRM for the full approach.

A churn-reduction checklist

Written definitions of activity, churn and dormancy, used in every report.
Head-count churn and value churn by cohort.
A per-player baseline gap for trigger rules.
A prediction target with a history window, an inactivity threshold and a horizon.
A time-based validation, a leakage check and a calibration check.
Suppression rules that run before every campaign.
A randomised holdout for each programme.
Results reported as incremental NGR after cost.
A monthly review of model drift and test outcomes.

CALCULATOR

Calculate returns for your player base

Pick a market, a campaign goal and the number of players. The calculator models NGR, ROMI and payback over 12 months.

Open the calculator

LIFECYCLE ASSESSMENT

Find where your CRM loses players

We review CRM, bonuses, loyalty, player journeys and channels, then give you a short improvement plan.

Order a diagnostic

Frequently asked questions

What is a good player churn rate in iGaming?

The number depends on the activity event, the window and the product mix, so a 30-day rate from one operator does not compare with a 90-day rate from another. Benchmark against your own earlier periods and your own cohorts, with the definition frozen.

How do you calculate player churn rate?

Fix a base of players who were active in a reference window, for example the last 30 days. Count how many of them have no qualifying activity in the next window. Churn rate is churned players divided by the base. Retention rate is one minus churn rate.

How long without activity makes a casino player churned?

It depends on the player's normal rhythm. A daily slots player silent for two weeks is a stronger signal than a monthly sports bettor silent for two weeks. Use a fixed window of 30, 60 or 90 days for reporting and a player-specific gap for triggering messages.

Can machine learning predict player churn?

Yes, well enough to rank players. A published study on one Belgian online gambling operator used recurrent neural networks and reached precision of 0.62 to 0.64 and recall of 0.49 to 0.61 for the churn class, depending on the architecture. That ranks players for a campaign. It does not make any single prediction certain.

How do you know a retention campaign reduced churn?

Hold out a randomly selected control group from the same eligible audience and send it nothing. Compare deposits and NGR per player between the treated and control groups over the same period. The difference is the effect of the campaign.

Should we contact every player who is likely to churn?

No. Remove self-excluded players, players using time-outs or cooling-off tools and anyone showing strong indicators of harm. Respect channel consent. A churn score is a commercial signal and never overrides a safer gambling decision.