ML for Retention
What is per-player scoring?
Per-player scoring is a retention approach where a model returns one decision per individual player, instead of assigning everyone in a cohort the same treatment.
The short version
Cohort retention groups players by a few observable traits - deposit tier, days since last session, country - and treats everyone inside the group identically. Per-player scoring drops the grouping. A model reads that player's own transaction history and returns a number for that player alone: churn risk, reactivation probability, or expected value of contact.
Why cohorts leak money
Inside any cohort, the behaviour spread is enormous. A cohort of 'deposited in the last 30 days, no session in 7 days' contains a VIP whose bet size quietly halved and a bonus hunter who was never going to return. Send them the same offer and you overpay one and underserve the other. The cost is invisible because the cohort average still looks fine.
What the model actually reads
Raw transactions, not aggregates. Deposit and withdrawal sequence with timestamps and amounts. Session cadence and duration. Bet-size trajectory over rolling windows. Game-mix shifts. Bonus acceptance and wagering completion. Support contacts. The features that matter are almost always about change over time, not level at a point in time.
Why the output has to be explainable
A score no analyst can interrogate does not get used. Per-player scoring should return the features that drove the flag alongside the number - 'bet size down 62% over 14 days, session gap 3x the personal baseline'. That is what turns a model output into an action a retention team is willing to take.
Where this fits
Per-player scoring makes sense once you have roughly six months of player history and a team that already runs retention manually. It upgrades prioritisation for a team that exists. It does not replace the team, and it does not fix a product that players are leaving for product reasons.
We build automation systems for small teams. Free audit call to map your specific workflows - no pitch, just a plan.