churn-risk-detection

# Churn Risk Detection ## Overview This skill identifies customers exhibiting churn signals by analyzing purchase recency decay, frequency decline, engagement drop-off, and behavioral anomalies. It produces a risk-scored customer list with churn probability estimates, time-to-churn predictions, revenue-at-risk quantification, and prescribed intervention strategies. Designed for non-contractual CPG and retail contexts where churn is silent (customers simply stop buying). ## When to Use - Proactively identifying customers likely to churn before they lapse - Building retention trigger campaigns based on behavioral signals - Quantifying revenue at risk from customer attrition - Evaluating the health of a customer base over time - Designing and prioritizing win-back campaigns - Setting up automated churn prevention workflows in CRM/ESP ## Required Inputs | Input | Required | Description | |---|---|---| | Transaction Data | Yes | Customer-level purchase history with dates, amounts, and order counts | | Customer IDs | Yes | Unique identifiers for each customer | | Analysis Date | Yes | Current date or reference date for recency calculations | | Category Purchase Cycle | Recommended | Expected repurchase interval for product category (e.g., 30 days for coffee) | | Engagement Data | Recommended | Email opens/clicks, site visits, app sessions over time | | Subscription Status | No | Active, paused, cancelled subscription status (if applicable) | | Customer Service Data | No | Support tickets, complaints, returns, refund history | | Acquisition Channel | No | How the customer was acquired (organic, paid, referral, etc.) | ## Methodology ### Step 1 — Define Churn for the Business In non-contractual settings, churn must be operationally defined: **Category-Based Churn Definition**: ``` Churn Threshold = Expected Repurchase Cycle × Churn Multiplier Recommended Multipliers: Consumables (coffee, supplements, cleaning): 2.0× → e.g., 30-day cycle → churn at 60 days Semi-durable (skincare, personal care): 2.5× → e.g., 60-day cycle → churn at 150 days Durable (cookware, appliances): 3.0× → e.g., 180-day cycle → churn at 540 days Grocery basket: 1.5× → e.g., 14-day cycle → churn at 21 days ``` If category cycle is unknown, calculate from data: ``` Median Inter-Purchase Interval (IPI) = Median of (Order Date N+1 - Order Date N) across all customers Churn Threshold = Median IPI × 2.0 (or use 75th percentile IPI × 1.5) ``` ### Step 2 — Behavioral Signal Extraction Extract churn signals across multiple dimensions: **Purchase Behavior Signals**: | Signal | Calculation | Churn Indicator | |---|---|---| | Recency Gap | Days since last purchase ÷ Avg IPI | Ratio >1.5 indicates concern; >2.0 critical | | Frequency Decline | Orders in last 90d vs. prior 90d | >30% decline is a strong churn signal | | AOV Decline | AOV last 3 orders vs. lifetime AOV | >20% decline indicates reduced commitment | | Basket Shrinkage | Items per order trending down | Reducing engagement with product range | | Category Narrowing | # categories last 3 orders vs. historical | Down-trading to fewer categories | | Promotion Dependency | % of recent orders with discount code | Increasing to >80% signals price-only loyalty | **Engagement Signals**: | Signal | Calculation | Churn Indicator | |---|---|---| | Email Open Decay | 30-day open rate vs. 90-day average | >40% decline | | Click-through Decline | 30-day CTR vs. 90-day average | >50% decline | | Site Visit Frequency | Sessions last 30d vs. prior 30d | >50% decline | | App Uninstall | App removed or sessions dropped to zero | Strong churn signal | | Loyalty Inactivity | Points earned last 60d = 0 (for loyalty members) | Disengagement signal | **Service Signals**: | Signal | Churn Indicator | |---|---| | Recent complaint (unresolved) | 2× churn risk elevation | | Multiple returns (3+ in 90 days) | 3× churn risk elevation | | Negative review or rating | 1.5× churn risk elevation | | Subscription downgrade or pause | Immediate intervention needed | | Delivery failure or late shipment | 1.5× churn risk elevation (compounding) | ### Step 3 — Churn Risk Scoring Model Build a composite churn risk score (0–100): **Weighted Signal Model**: ``` Churn Risk Score = (Recency Gap Score × 0.30) + (Frequency Decline Score × 0.25) + (Engagement Decay Score × 0.20) + (AOV/Basket Decline Score × 0.10) + (Service Issue Score × 0.10) + (Promotion Dependency Score × 0.05) Each component scored 0–100: Recency Gap Score = min(100, (Days Since Purchase ÷ Churn Threshold) × 100) Frequency Decline Score = min(100, max(0, (1 - Recent Freq ÷ Historical Freq) × 100)) ...etc. ``` **Risk Tier Classification**: | Tier | Score Range | Estimated Churn Probability | Urgency | |---|---|---|---| | Low Risk | 0–25 | <10% | Monitor quarterly | | Moderate Risk | 26–50 | 10%–30% | Monitor monthly; soft engagement | | High Risk | 51–75 | 30%–60% | Active intervention within 2 weeks | | Critical Risk | 76–100 | >60% | Immediate intervention; likely churning now | ### Step 4 — Revenue-at-Risk Quantification Calculate the financial impact of potential churn: ``` Revenue at Risk (per customer) = Predicted Annual Revenue × Churn Probability Where Predicted Annual Revenue = Historical AOV × Historical Annual Frequency Aggregate Revenue at Risk = Σ (Revenue at Risk per customer) for all customers with score >50 ``` **Segment-Level Risk Summary**: ``` Segment | Customers at Risk | Avg Churn Score | Revenue at Risk | % of Total Revenue Champions | XX | XX | $XX,XXX | X% Loyal | XXX | XX | $XXX,XXX | XX% At Risk | X,XXX | XX | $XXX,XXX | XX% Total | X,XXX | -- | $X,XXX,XXX | XX% ``` ### Step 5 — Churn Driver Analysis Identify the primary churn drivers in the customer base: 1. **Recency-driven churn**: Customers simply haven't returned — no engagement decline, just inactivity 2. **Experience-driven churn**: Customers had negative experiences (returns, complaints, delivery issues) 3. **Value-driven churn**: Customers only buy on promotion; churn when discounts stop 4. **Competition-driven churn**: Cross-shopping signals (declining share of wallet) 5. **Lifecycle-driven churn**: Natural category exit (e.g., baby products as children age out) 6. **Subscription fatigue**: Active subscription cancellations or skip frequency increasing Map each at-risk customer to their primary churn driver for targeted intervention. ### Step 6 — Intervention Prescription Prescribe actions based on risk tier and churn driver: | Risk Tier | Driver | Intervention | Channel | Timing | |---|---|---|---|---| | High | Recency | "We miss you" + incentive | Email + SMS | Day 1 of high-risk classification | | High | Experience | Service recovery outreach | Personal email or phone | Within 48 hours | | High | Value | Exclusive loyalty offer (not discount) | Email | Day 3 | | Critical | Recency | Escalated offer + free shipping | SMS + push | Immediate | | Critical | Subscription | Pause option + downsell offer | Email + in-app | Pre-cancellation trigger | | Moderate | Recency | Content re-engagement (new products, tips) | Email | Weekly cadence | | Moderate | Value | Bundle offer or subscribe-and-save pitch | Email | Next promotional window | **Incentive Ladder** (escalating offers for non-responsive at-risk customers): ``` Day 0: Personalized product recommendation (no incentive) Day 7: 10% off next order Day 14: 15% off + free shipping Day 21: 20% off + free gift with purchase Day 30: Final win-back: 25% off "last chance" offer Day 45: Move to suppression list; reduce marketing spend ``` ### Step 7 — Monitoring & Alert System Design Define ongoing churn monitoring: - **Daily scan**: Flag newly critical-risk customers for immediate action - **Weekly digest**: Summary of risk tier migration (how many moved from moderate to high?) - **Monthly review**: Churn rate trend, intervention effectiveness, revenue-at-risk dashboard - **Quarterly recalibration**: Adjust scoring weights based on observed churn vs. predicted churn **Alert Triggers**: - Customer crosses from moderate to high risk → trigger retention workflow - High-value customer (top 10% LTV) enters high risk → alert customer success team - Churn rate exceeds baseline by >20% → flag systemic issue for investigation - Subscription cancellation initiated → trigger save flow ## Output Specification 1. **Churn Risk Scorecard**: Every customer scored with risk tier, probability, and primary driver 2. **Revenue-at-Risk Summary**: Aggregate and segment-level revenue exposure 3. **Top 50 At-Risk Customers**: Prioritized list of highest-value customers at greatest risk 4. **Churn Driver Distribution**: Breakdown of primary churn causes across the at-risk population 5. **Intervention Playbook**: Per-tier, per-driver recommended actions with channel and timing 6. **Monitoring Dashboard Spec**: Metrics, thresholds, and alert rules for ongoing churn tracking ## Examples **Input**: "Identify churn risk for our pet food DTC brand. 25,000 customers, average repurchase cycle is 28 days. We've noticed a spike in subscription cancellations." **Output**: 3,200 customers (13%) classified as high or critical risk, representing $420K in annual revenue at risk. Subscription cancellers (800 customers) are the highest-risk cohort; primary driver is subscription fatigue (average tenure 8 months). Recommendation: introduce "pause" option, flexible delivery frequency, and surprise-and-delight program at month 6. Non-subscription at-risk customers show recency-driven patterns; prescribe a 4-step incentive ladder. **Input**: "Our beauty brand has 60% first-year churn. Help us understand why and who's most at risk." **Output**: Churn analysis reveals 42% of first-year churn happens before the 2nd purchase (within 60 days). Primary driver: post-purchase disengagement (no email engagement after order confirmation). High-risk new customers identified by: no email open within 14 days, no site revisit within 30 days, and single-SKU first order. Intervention: redesigned post-purchase nurture sequence with education content, usage tips, and day-45 reorder incentive. ## Guidelines - Non-contractual churn (CPG/retail) requires probabilistic estimation — there is no single "churn event" - Always define churn threshold relative to category purchase cycle, not arbitrary time periods - Combine behavioral signals; no single metric reliably predicts churn alone - High-value customers warrant more aggressive (and more expensive) retention interventions - For low-value, high-churn segments, it may be more efficient to let them churn than to invest in retention - Distinguish between addressable churn (can be prevented) and structural churn (lifecycle exit) - Track intervention effectiveness: what % of high-risk customers were retained after intervention? - Avoid "discount addiction" — escalate non-monetary value (exclusive access, content, community) before discounts - For subscription businesses, monitor skip rate and frequency changes as early warning signals ## Validation Checklist - [ ] Churn is operationally defined with category-appropriate thresholds - [ ] Multiple behavioral signals are combined (not relying on recency alone) - [ ] Risk scores are calibrated against observed churn rates - [ ] Revenue at risk is quantified at customer and segment level - [ ] Churn drivers are identified and mapped to specific interventions - [ ] High-value at-risk customers are prioritized for immediate action - [ ] Intervention playbook includes escalation ladder and channel recommendations - [ ] Monitoring system includes automated alert triggers - [ ] False positive rate is assessed (customers flagged but didn't actually churn)