Health Score Inputs for DTC: RFM + Support + Payments

Q: What's the minimum sample size to validate a health score model?

Minimum 500 churned customers with 12 months of prior behavior data. Smaller brands should pool 18 - 24 months of churn history. Run logistic regression to confirm each input's predictive power (p < 0.05). If support or payment data is sparse, increase sample size to 1,000+.

Why RFM Alone Misses Churn Signals

Recency, Frequency, Monetary (RFM) scoring captures purchase momentum but ignores operational friction. A customer with strong RFM metrics who files three support tickets per month or has recurring payment declines is at higher churn risk than raw transaction data suggests.

DTC brands see 20 - 40% of churn driven by support experience and payment friction, not purchase inactivity. Health scoring must weight these signals equally with purchase behavior to avoid false negatives in retention modeling.

RFM Component Thresholds

Establish baseline thresholds for each RFM dimension based on brand AOV, repeat rate, and category. Recalculate quarterly to account for seasonality and cohort drift.

Recency: Days since last purchase. Threshold varies by category (e.g., 60 days for CPG, 180 days for apparel). Score 0 if > 2x median repeat cycle.
Frequency: Purchases in trailing 12 months. Benchmark: one-time buyers score 20; 2 - 3 purchases score 60; 4+ purchases score 90.
Monetary: Total spend in trailing 12 months or LTV. Segment into quartiles. Bottom quartile = 20 points; top quartile = 90 points.

Support Friction Signals

Support interactions reveal product - market fit and operational quality. High ticket volume relative to order count signals product confusion, quality issues, or unclear communication. Track both volume and resolution velocity.

Establish a support ratio: support tickets per 100 orders. Benchmark: 2 - 5 tickets per 100 orders is healthy; 10+ per 100 is elevated risk. Tickets unresolved after 7 days reduce health score by 15 points per ticket.

Ticket volume: Count all support interactions (email, chat, phone) in trailing 90 days. Divide by order count in same period. Multiply by 100.
Resolution time: Median days from ticket open to resolution. Threshold: < 2 days = 80 points; 2 - 5 days = 50 points; > 5 days = 20 points.
Repeat issues: Same customer filing 3+ tickets on same topic in 90 days = 30 - point penalty. Indicates unresolved root cause.
Sentiment: If available, flag negative language in tickets. Each negative ticket = 10 - point deduction.

Payment Reliability Indicators

Payment failures and disputes are direct churn predictors. A customer with one failed payment attempt is 2x more likely to churn within 90 days. Track both hard failures (declined card, insufficient funds) and soft signals (multiple retry attempts).

Subscription brands must monitor dunning cycles. Customers requiring 3+ dunning attempts have 60% churn probability within 30 days of recovery.

Failed transactions: Count declined payments in trailing 90 days. First failure = 20 - point deduction; each additional = 15 points.
Dunning attempts: For subscriptions, count retry cycles. 1 cycle = 10 - point deduction; 2+ cycles = 40 - point deduction.
Dispute rate: Chargebacks or disputes as % of total transactions. Threshold: > 0.5% = 25 - point penalty.
Payment method changes: Frequent updates (3+ in 90 days) suggest card issues or account instability. 15 - point deduction per change.

Composite Scoring Formula

Combine all inputs into a weighted score. Adjust weights based on brand cohort analysis - test different allocations against actual churn data to validate.

Health Score = (RFM Score × 0.50) + (Support Score × 0.25) + (Payment Score × 0.25)
RFM Score: Average of Recency, Frequency, Monetary (each 0 - 100).
Support Score: 100 - (Ticket Ratio Penalty + Resolution Time Penalty + Repeat Issue Penalty). Floor at 0.
Payment Score: 100 - (Failed Transaction Penalty + Dunning Penalty + Dispute Penalty). Floor at 0.
Recalculate weekly. Flag customers with score < 40 for retention outreach within 48 hours.

Segmentation and Action Triggers

Use health score bands to route customers to specific workflows. Avoid one - size - fits - all retention messaging.

Score 70 - 100 (Healthy): Standard engagement. Monitor for upsell signals.
Score 40 - 69 (At Risk): Trigger win - back email series. Offer discount or exclusive product. Assign to support for proactive outreach if score < 50.
Score < 40 (Critical): Immediate intervention. Direct SMS or phone outreach. Investigate support and payment issues. Offer concession (refund, replacement, credit).
Score increase of 20+ points in 30 days: Customer recovered. Move to standard segment. Track what action drove recovery.

Monitoring and Calibration

Health scores are only useful if validated against actual churn. Run monthly cohort analysis to confirm that low - score customers churn at predicted rates. Adjust weights and thresholds quarterly based on model performance.

Track false positives (high score, churned anyway) and false negatives (low score, retained). Use these to refine input selection and weighting.

Churn validation: For customers who churned in past 30 days, calculate what their health score was 60 days prior. If median score of churned cohort is > 50, model is underweighting risk signals.
Cohort drift: Recalculate RFM thresholds quarterly. Seasonal spikes or new product launches shift baseline purchase patterns.
A / B test interventions: Segment at - risk customers randomly. Send retention offer to 50%; hold 50% as control. Measure incremental retention lift. Adjust outreach strategy based on results.

Questions

FAQ

What if a brand has no support ticket system?

Use proxy signals: count refund requests, return initiations, and customer service emails as tickets. If data is unavailable, reduce support weight to 10% and increase RFM weight to 60%. Implement a ticketing system within 60 days - support friction is invisible without it.

How often should health scores be recalculated?

Weekly minimum for weekly - cohort brands (subscription, CPG). Daily for high - frequency categories (e.g., food delivery). Recalculate thresholds and weights quarterly. Avoid daily recalculation for monthly - repeat categories - noise outweighs signal.

Should new customers have a different health score model?

Yes. Exclude Recency for customers < 90 days old (all are recent). Use Frequency and Monetary as leading indicators instead. Apply support and payment signals at full weight. Graduate to standard model at 90 days.

What's the minimum sample size to validate a health score model?

Minimum 500 churned customers with 12 months of prior behavior data. Smaller brands should pool 18 - 24 months of churn history. Run logistic regression to confirm each input's predictive power (p < 0.05). If support or payment data is sparse, increase sample size to 1,000+.

Want this on your account?

Thirty minutes. Bring the number that keeps you up.

Book a demo