Fighting Game · Guides & deep dives

V2 Churn Prediction Model

exposed

3sections1 minread
Markdown source
On this page

Phase 72.5.2.11 adds @v2/churn-prediction-model, the player churn prediction model for accounts likely to stop playing in the next 7 days.

The model composes @iris/analytics retention thresholds and uses buildV2ChurnPredictionModelSurface to transform hashed-account engagement telemetry into an interpretable logistic-regression score. It accepts v2.player.account.created, v2.player.session.started, v2.player.session.ended, v2.player.session.activity, and v2.cosmetic.purchased events. Raw account IDs are not accepted or emitted.

Policy#

  • Policy ID: v2-churn-prediction-model-v1
  • Source library: @iris/analytics
  • Model type: interpretable logistic regression
  • Prediction horizon: 7 days
  • Required output: churn probability, risk banding, feature driver attribution
  • Privacy: account-level joins use accountIdHash; raw account IDs are never exposed
  • Retention action guardrail: re-engagement must use cosmetic grants, free-weekend invites, or non-paywall reminders; paywall pressure is forbidden

Output#

The surface emits:

  • per-account churnProbability and low/medium/high/critical risk bands
  • likelyToChurnWithin7Days decisions using a default 0.65 risk threshold
  • feature values for days since last session, recent/previous session frequency, session-length drop, active days, purchases, and tenure
  • top feature drivers and ethical recommended action for each prediction
  • aggregate risk distribution and top at-risk player rows

Gates#

check-v2-churn-prediction-model.py validates the contract, service package, source tokens, tests, docs, workflow wiring, Horde gates, and the Phase 72 checklist item. Required gates are churn-prediction-model-service-present, churn-prediction-model-seven-day-horizon, churn-prediction-model-risk-scoring, churn-prediction-model-feature-drivers, and churn-prediction-model-ci-wired.

Targeted verification:

bash
pnpm --filter @v2/churn-prediction-model run typecheck
pnpm --filter @v2/churn-prediction-model run test
python V2/ue/Tools/check-v2-churn-prediction-model.py