Cohort Analysis for iGaming

The four choices every cohort hides

Choice	Options	How it moves the number
Anchor	Registration, first deposit, first bet	Registration cohorts include everyone who never funded; FTD cohorts look far healthier on the same players
Window	Player-relative (day 0–30 from each anchor) vs calendar (activity in a month)	Calendar windows mix cohort ages; player windows are comparable but mature later
Filters	Test, bot, fraud, self-excluded accounts in or out	Funnel and retention rates shift whenever the exclusion list changes — silently, unless the filter set is versioned
Censoring	Immature players included, excluded, or projected	Including them understates recent cohorts; the newest campaign always "underperforms"

None of these choices is wrong. What is wrong is leaving them implicit — because then every team makes them independently, and the monthly meeting spends its hour discovering the differences instead of reading the business.

A worked illustration (synthetic numbers)

Take an illustrative cohort of 1,000 registrations, of whom 400 deposit. Day-30 "retention" can honestly be reported as 12% (active players over all registrations), 30% (active over depositors), or 45% (active over depositors, whale-trimmed medians, mature subset only) — from the same underlying behaviour. All three are defensible; none is comparable to the others. The fix is not choosing the "right" one; it is naming the one you use and never switching mid-chart.

Reading a cohort table without fooling yourself

Read down the diagonal, not across the bottom row. The newest row is always the most censored; its tail is missing data, not decay.
Watch the denominators between rows. A retention "improvement" that coincides with an acquisition-mix change is usually the mix, not the product. Split by source before celebrating — media-buy and organic cohorts behave structurally differently.
Trim or median the value metrics. Averages over stake-distributed populations answer to a handful of accounts; decide the trimming rule once and keep it.
Anchor product claims to matched cohorts. "Players who used X retain better" is self-selection until the comparison group is matched or randomised — the recurring theme of every measurement page in this Academy.
Respect the second-deposit checkpoint. FTD-to-second-deposit matures fast and predicts the tail early; it is the honest early read while LTV horizons are still censored.

Where cohorts live in the stack

Cohort analysis is a consumer of everything upstream: the event schema supplies the anchors and both clocks, the warehouse supplies player-grain facts and versioned filters, and the dictionary pins the four choices so that "day-30 retention" means one thing in every deck. When a cohort number surprises you, debug in that order — schema, warehouse, definition — before believing the behaviour changed.

The reporting contract

Any cohort figure that leaves the analytics team should carry its four choices inline: anchor, window, filters, censoring treatment — one line of fine print that costs nothing and ends most disputes before they start. A number without that line is an invitation for someone to reproduce it differently and call yours wrong.

Continue reading: Dashboards by role — where these reads get consumed daily. Cohort retention — the glossary definition this page operationalises.

Cohort analysis — common questions

What defines a cohort properly?

Four fields, all explicit: the anchor event (registration? first deposit? — they produce different populations), the anchor window (which dates enter), the eligibility filters (test, fraud, self-excluded accounts in or out), and the observation window measured from each player's own anchor, not from a calendar date. Any cohort claim missing one of the four is under-specified, and two teams will legitimately compute different values from it.

Why do retention numbers disagree between teams?

Usually one of three silent choices: a different anchor (registration versus FTD cohorts can differ dramatically), calendar-window versus player-window observation, or different treatment of immature cohorts — including players whose window has not finished yet drags the tail down. All three are legitimate methods; mixing them across teams is what produces the disagreement. Pin the choices in the metric dictionary.

What is censoring and why does it distort LTV?

A player whose observation window has not yet completed is 'censored' — you do not know their final value, only their value so far. Averaging censored players in as if they were finished systematically understates recent cohorts and makes every new campaign look worse than the old ones. The honest treatments are to report only mature cohorts at each horizon, or to state the censoring rule explicitly when projecting. See LTV for the definitional side.

When is a cohort big enough to read?

When the movement you are reading is larger than the movement the cohort produces by luck. Small cohorts swing hard — a handful of high-value players can move an average by a lot, which is why whale-filtering or medians belong in the method, not in the appendix. If a segment is too small to read honestly, the answer is a longer window or a coarser segment, not a confident narrative over noise.

How do cohorts connect to A/B tests and holdouts?

A cohort read describes what happened to a defined population; only a comparison against a control tells you what caused it. The pattern that survives scrutiny is cohort + matched control + pre-registered window — the same discipline our personalization case study uses, and the reason bonus economics insists on holdouts before ROI claims.

Cohorts, honestly.