Churn Prevention Scoring

Identifies customers at risk of churning 30-60 days before traditional signals appear. Proven to reduce churn by 40% in production (12% → 7.2% monthly churn in financial services deployment).

How It Works

Customer Activity Data → Churn Feature Engineering → Survival Model + XGBoost
    → 30/60/90-Day Risk Scores → CDP Retention Segments → Automated Outreach

Churn Definition Framework

Before building the model, define what "churn" means for your business:

Business Type	Churn Definition	Observation Window
Subscription	Cancellation or non-renewal	30 days post-renewal date
Transactional	No purchase in 2x average purchase cycle	Rolling 90-day window
Platform	Zero logins for 30+ days after being active	Rolling 30-day window
Hybrid	Composite: revenue drop >80% AND engagement drop >70%	60-day window

Feature Engineering

Leading Indicators (pre-churn signals)

-- Engagement decay: the single most predictive churn feature
SELECT
  customer_id,
  -- Engagement velocity (negative = declining)
  SAFE_DIVIDE(
    COUNTIF(event_date >= DATE_SUB(CURRENT_DATE(), INTERVAL 7 DAY)),
    COUNTIF(event_date BETWEEN
      DATE_SUB(CURRENT_DATE(), INTERVAL 30 DAY)
      AND DATE_SUB(CURRENT_DATE(), INTERVAL 7 DAY))
  ) - 1.0 AS engagement_velocity,

  -- Support contact acceleration
  COUNTIF(
    event_type = 'support_contact'
    AND event_date >= DATE_SUB(CURRENT_DATE(), INTERVAL 14 DAY)
  ) AS recent_support_contacts,

  -- Payment failure signals
  COUNTIF(
    event_type = 'payment_failure'
    AND event_date >= DATE_SUB(CURRENT_DATE(), INTERVAL 30 DAY)
  ) AS payment_failures_30d,

  -- Feature usage breadth decline
  COUNT(DISTINCT CASE
    WHEN event_date >= DATE_SUB(CURRENT_DATE(), INTERVAL 30 DAY)
    THEN feature_used END) AS features_used_30d,
  COUNT(DISTINCT CASE
    WHEN event_date BETWEEN
      DATE_SUB(CURRENT_DATE(), INTERVAL 60 DAY)
      AND DATE_SUB(CURRENT_DATE(), INTERVAL 30 DAY)
    THEN feature_used END) AS features_used_prior_30d

FROM `project.analytics.customer_events`
GROUP BY customer_id

Top Churn Predictors (from production models)

1. Engagement velocity — Rate of change in activity over time
2. Support contact frequency — Increasing contacts = friction signal
3. Feature usage contraction — Using fewer features than before
4. Payment failures — Hard churn predictor
5. Competitive exposure — Visits to competitor sites (if trackable)
6. NPS/CSAT decline — Survey score drop from prior period
7. Time-to-value stall — Not reaching activation milestones

Multi-Horizon Scoring

ConversionOS generates risk scores at three time horizons:

Horizon	Model	Primary Action
30-day risk	XGBoost classifier	Trigger automated retention journey
60-day risk	Survival analysis (Cox PH)	Flag for personal outreach
90-day risk	Ensemble (XGBoost + survival)	Adjust LTV forecasts, inform acquisition spend

Retention Campaign Integration

{
  "retention_triggers": {
    "tier_1_critical": {
      "condition": "churn_30d_score >= 0.8 AND ltv_tier = 'HIGH'",
      "action": "personal_outreach",
      "channel": "phone_call",
      "sla": "24_hours"
    },
    "tier_2_elevated": {
      "condition": "churn_30d_score >= 0.6",
      "action": "automated_journey",
      "channel": "email_sms",
      "journey": "retention_winback_v3"
    },
    "tier_3_monitoring": {
      "condition": "churn_60d_score >= 0.5 AND churn_30d_score < 0.6",
      "action": "add_to_nurture",
      "channel": "email",
      "journey": "engagement_boost"
    }
  }
}

Production Results

From financial services deployment (anonymized):

Metric	Before	After	Impact
Monthly churn rate	12.0%	7.2%	-40%
At-risk detection lead time	Reactive (post-churn)	30-day predictive	+30 days
Retention campaign ROI	1.2x	3.8x	+217%
Annual revenue preserved	—	$4.2M	—