Senior operator playbook on cohort retention modeling for online casino. D1/D7/D30/D90/D180 curves, NGR per cohort, BG/NBD vs Pareto/NBD, payback by channel.
Cohort Retention Modeling for Casino Operators 2026 This guide is written for the operator-side head of CRM , head of analytics, or VP of finance who owns cohort economics for an online casino brand. It is not a "what is retention" guide. It assumes you already have a CRM platform, a player database with deposit and wager records, and the question is: how do I model cohort retention with the precision needed to drive channel allocation, bonus engineering, and CRM investment decisions.
Why cohort modeling beats aggregate retention Aggregate retention metrics (overall active players this month vs last month) hide three problems that destroy operator economics:
Channel mix change masks retention change. An operator that shifts acquisition from low-retention channels to high-retention channels will show aggregate retention improvement that has nothing to do with retention work and everything to do with channel mix. The reverse is also true.Cohort decay shape matters more than cohort size. Two cohorts of the same size can produce dramatically different LTV depending on the decay curve shape (steep early decline with a long tail vs gentle decline that flattens later). Aggregate metrics cannot see this.CRM intervention effects are invisible at the aggregate. A CRM journey rollout that improves retention for the affected cohort produces a sub-1-percent aggregate retention shift that the CFO will not credit to CRM. Cohort modeling makes the intervention visible.Operators that run on aggregate retention metrics make channel allocation, bonus, and CRM decisions on noisy data. Operators that run on cohort modeling make those decisions on signal.
The retention curve: D1 to D180+ The core casino cohort retention curve tracks the share of a depositor cohort that is still active (defined as having wagered at least once in a defined window) at days 1, 7, 30, 90, and 180 after the FTD date.
The qualitative shape of a healthy retention-focused casino cohort:
D1 retention reflects welcome-bonus engagement and is typically the highest active-share point in the curve D7 retention shows the first reactivation flow effect D30 retention captures the first major dropoff D90 retention reveals the true engaged cohort emerging after bonus burn D180 retention is the long-tail engaged cohort For acquisition-focused brands (operators with aggressive bonus economics and less CRM discipline) the curve compresses meaningfully across all windows. The difference between retention-focused and acquisition-focused brands at D180 is the difference between sustainable operator economics and a treadmill where acquisition must run faster every quarter to offset retention bleed.
Specific benchmarks vary widely by jurisdiction, vertical mix, payment-method profile and product. Operators should baseline against their own historical cohorts rather than imported numbers from another market or operator.
NGR per cohort, not GGR per cohort GGR (Gross Gaming Revenue , the operator's win before bonus and other costs) is the wrong metric for cohort economics. NGR (Net Gaming Revenue , GGR minus bonus cost, free bets, comp credits) is the right metric because it reflects the actual revenue that the cohort generates after the acquisition incentives have been accounted for.
Track NGR per cohort at the same milestones as retention: D1, D7, D30, D90, D180, D365. The shape of the NGR curve is more informative than the absolute number because it reveals how the bonus economics decay relative to player engagement.
A healthy retention-focused brand sees NGR per cohort climb steadily as bonus liability burns down and engaged player NGR contribution compounds. An acquisition-focused brand sees NGR per cohort that stays low or even negative through the early window (bonus burn exceeds wager NGR) and only begins to climb once bonuses are exhausted.
Payback period: the channel allocation metric Payback period is the number of days or months from FTD until the cumulative cohort NGR equals the cohort acquisition cost (CPA times cohort size). It is the single most important channel allocation metric because it directly answers: how long is this operator's capital locked up per dollar of acquisition?
The qualitative ranking of channels by typical payback for a retention-focused casino brand:
Brand search and direct: shortest payback (low CPA , high quality cohort) SEO organic: short payback (no CPA in the classic sense, but content investment amortizes) Affiliate revshare partnerships: medium payback (no upfront CPA but ongoing revshare cost compresses NGR per cohort) Affiliate CPA partnerships: medium payback (upfront CPA load with variable cohort quality) Meta and Google paid: medium-long payback (volume engines with disciplined creative governance) Programmatic display: long payback (lower quality cohort with longer payback) Influencer paid (CPA -based): long payback (variable cohort quality and CPA economics) For acquisition-focused brands the payback periods extend across the board because the cohort NGR shape is shallower.
Operators that segment payback by channel and adjust channel allocation quarterly outperform operators that run a blended CPA target across all channels. The blended-CPA approach overpays for low-quality channels and underpays for high-quality channels.
BG/NBD vs Pareto/NBD: model selection for casino LTV Casino LTV modeling sits in the same family of probabilistic customer behavior models. Two foundational models dominate: BG/NBD (Beta-Geometric / Negative Binomial Distribution, Fader and Hardie 2005) and Pareto/NBD (Schmittlein, Morrison, Colombo 1987).
The differences matter for casino in specific ways:
BG/NBD assumes a customer becomes inactive immediately after a purchase (in casino terms, after a wager session) with some probability. The customer is alive while they keep playing; they die instantly between sessions with a churn probability.
Pareto/NBD assumes a customer's lifetime is exponentially distributed independent of purchase events. The customer is alive for some unobserved lifetime, during which they purchase at a Poisson rate, then dies.
For casino with our cohort retention shape (sharp early dropoff, long tail), BG/NBD typically fits the bulk of the cohort better but underfits the long-tail engaged players. Pareto/NBD typically fits the long-tail engaged players better but underfits the early-dropoff cohort.
Operator-grade approach for 2026: fit both models, use BG/NBD for the bulk cohort LTV projection (which drives the channel allocation decision), and use Pareto/NBD for the VIP and high-LTV tail segmentation (which drives the VIP host program and the high-value reactivation work).
Integration with bonus engine economics Bonus engine economics directly drive cohort retention and LTV. The cohort retention model has to integrate with the bonus model or both produce misleading numbers.
The integration points:
Bonus cost amortization across the cohort. A welcome bonus given on FTD is acquisition cost (CPA contributor), not retention cost. A reload bonus given on D30 is retention cost (NGR reducer). The model has to allocate bonus cost to the right cohort timeline window.Wagering requirement and bonus burn timing.Bonus-aided retention vs organic retention. Some D30 retention is driven by reload bonuses; some is driven by organic engagement. Operators that cannot distinguish the two end up paying for reload bonuses that would have driven retention without them. A/B testing reload bonus assignment within a cohort produces the lift estimate that should inform reload bonus policy.Cohort modeling in SQL and dbt Most operator data stacks in 2026 run on a warehouse (Snowflake, BigQuery, Redshift) with dbt as the transformation layer. The cohort retention model is built as a set of dbt models on top of the deposit and wager fact tables.
Core dbt models for cohort retention:
```sql -- dbt: fct_cohort_first_deposit.sql -- One row per depositor with the FTD cohort assignment select player_id, min(deposit_at) as ftd_at, date_trunc('month', min(deposit_at)) as ftd_cohort_month, date_trunc('week', min(deposit_at)) as ftd_cohort_week, -- carry through the acquisition attribution first_value(acquisition_channel) over ( partition by player_id order by deposit_at ) as acquisition_channel from {{ ref('fct_deposits') }} group by player_id; ```
```sql -- dbt: fct_cohort_retention.sql -- Player-day grain: was the player active on a given day after FTD with cohort as ( select * from {{ ref('fct_cohort_first_deposit') }} ), wager_activity as ( select player_id, date_trunc('day', wager_at) as activity_day, sum(wager_amount) as daily_wager, sum(wager_amount - payout_amount) as daily_ggr, sum(bonus_cost) as daily_bonus_cost from {{ ref('fct_wagers') }} group by 1, 2 ) select c.player_id, c.acquisition_channel, c.ftd_cohort_month, c.ftd_cohort_week, c.ftd_at, datediff('day', c.ftd_at, w.activity_day) as days_since_ftd, w.daily_wager, w.daily_ggr, w.daily_bonus_cost, (w.daily_ggr - w.daily_bonus_cost) as daily_ngr from cohort c left join wager_activity w on w.player_id = c.player_id and w.activity_day >= c.ftd_at; ```
```sql -- dbt: mart_cohort_retention_curves.sql -- Aggregate to cohort-month, days_since_ftd grain for the retention curve select ftd_cohort_month, acquisition_channel, days_since_ftd, count(distinct player_id) filter (where daily_wager > 0) as active_players, sum(daily_ngr) as cohort_ngr, count(distinct player_id) filter (where daily_wager > 0)::float / nullif( max(count(distinct player_id) filter (where days_since_ftd = 0)) over (partition by ftd_cohort_month, acquisition_channel), 0 ) as retention_rate from {{ ref('fct_cohort_retention') }} where days_since_ftd between 0 and 365 group by ftd_cohort_month, acquisition_channel, days_since_ftd; ```
These three dbt models, materialized incrementally and refreshed daily, give the analytics team the cohort retention curve, NGR per cohort, and channel-segmented retention rates that drive the channel allocation, bonus, and CRM decisions.
Layer the BG/NBD and gamma-gamma models on top using the `lifetimes` Python package or equivalent R packages. The probabilistic models consume the cohort fact table and produce LTV projections that the dbt warehouse stores as a separate mart.
Cohort decay shape across markets Retention-focused European casino brands typically show a more gradual decay shape across the D7 to D180 window than aggressive-promo LATAM brands; lower-ARPU higher-volume markets (Philippines e-Bingo, for example) show faster early decay but larger absolute cohort sizes that produce comparable cumulative NGR over the long window.
The CRM playbook differs by market (shorter reactivation intervals for high-velocity markets, payment-method-aware journey design where e-wallet penetration is high, fixture-anchored journeys where football or basketball drives the sports cadence) but the cohort math discipline is the same. Operators should always baseline against their own historical data rather than imported benchmarks; the value of cohort modeling lies in the operator-specific insights it surfaces.
How to use cohort models: three operational decisions The cohort model is only useful if it drives operational decisions. The three decisions it should drive:
Channel allocation. Quarterly rebalancing of acquisition spend across channels based on segmented payback periods. Cap or pause channels with payback periods materially worse than the operator's target; expand channels with payback periods materially better.Bonus and offer engineering. A/B testing of welcome bonus structure, reload bonus cadence, and wagering requirements with cohort-level retention curve comparison. Pick the variants that produce flatter D30-to-D90 decay shapes, not the variants that maximize D1 to D7 engagement (which often comes from heavier bonus loads that produce worse long-tail economics).CRM intervention prioritization. Identifying the cohort segments with the worst D30 to D90 decay and routing CRM journey investment there. A several-point improvement in D30 retention for a poor-performing segment delivers more LTV uplift than a smaller improvement in D90 retention for a well-performing segment.Operators that run the cohort model and then ignore its output for political or organizational reasons are common. Operators that build the cohort model into the quarterly planning rhythm of the CRM, acquisition, and product teams compound retention discipline into structural cohort economics advantage.
How Basher applies cohort modeling Cohort modeling work intersects three of Basher's eight services: managed CRM execution (the journey design and reactivation work that improves the decay shape), benchmarking and analytics (the modeling itself), and consulting (the channel allocation and bonus engineering decisions that the model drives). We run these alongside our acquisition services (traffic generation, media buying) so the cohort insight feeds back into the channel allocation it informs.
We see this play out across the Basher client and partner base: operators like Stake, Bet365 and Betano have the operator scale to support the full BG/NBD and Pareto/NBD layered approach; mid-tier operators like BetBoom, PIN-UP and Jackpoty benefit most from disciplined channel-segmented payback analysis where their CRM resourcing is concentrated.
Get the senior view If you are building cohort modeling capability, rebaselining underperforming cohort economics, or layering managed CRM execution on top of your existing platform, we can help.
