clv-modeling

CLV Modeling

Probabilistic customer lifetime value: BG/NBD for purchase frequency, Gamma-Gamma for monetary value, Beta-Geometric for subscriptions — with confidence intervals, CLV:CAC, cohort curves, and at-risk high-value alerts.

Contract

Role: Advisory modeler. This skill computes and reports predicted value; it does not target, price, or contact customers. All numbers come from deterministic Python scripts, never from model estimation in prose.

Mode:

• quick — RFM summary + single-horizon MLE point estimates.
• standard (default) — MLE BG/NBD + Gamma-Gamma, CIs, tiers, at-risk list, holdout validation.
• deep — add Bayesian (PyMC-Marketing) posteriors and full uncertainty.

When to use: predicting future per-customer revenue, CLV:CAC, at-risk high-value retention targets, purchase-frequency / probability-alive scoring.

When NOT to use: for building customer segments or cohort retention curves use audience-segmentation; for sales-lead propensity use crm-lead-scoring; for channel ROI use attribution-analysis. See ../../references/skill-index.md for the full portfolio map.

Evidence required (inputs):

• workspace/raw/transactions.csv — customer_id, date, amount (required for non-contractual). If absent, STOP and run data-extraction first.
• workspace/raw/subscriptions.csv — customer_id, start_date, end_date, plan_value (optional; enables contractual model).
• workspace/analysis/acquisition_costs.json — CAC by customer/segment (optional; enables CLV:CAC).

Depends on: data-extraction (raw transactions). Feeds into: audience-segmentation (value tiers), paid-media (bid targets), email-analytics (lifecycle triggers), attribution-analysis, reporting. Builder detail lives in references/authoring-notes.md.

Hard STOP (data integrity): stop and surface the issue before modeling if any of these appear — future-dated transactions, negative/zero amounts as a material share, or fewer than ~2 repeat purchasers (BG/NBD is unidentifiable). Do not emit CLV numbers from broken inputs.

Parse $ARGUMENTS for inline paths, horizon, discount rate, or model overrides.

Workflow

1. Validate & profile. Load transactions; parse dates robustly (ISO/US/EU). Flag duplicates (same customer/date/amount), negative/zero amounts, and impossible dates. Report customer count, date range, transaction volume. Apply the Hard STOP gate above.

2. RFM summary. Run scripts/rfm_summary.py to build recency, frequency (repeat purchases only), tenure (T), and mean repeat monetary value. One-time purchasers: frequency=0, monetary=NaN (excluded from Gamma-Gamma).

3. Model-class gate (AskUserQuestion). When both transaction and subscription data exist, or the business model is ambiguous, ask before fitting:

   • Question: "Which lifetime model fits this business?"
   • Options: (a) Non-contractual (BG/NBD + Gamma-Gamma) — purchases happen at will, no fixed renewal; (b) Contractual (Beta-Geometric) — subscription with renewal periods; (c) Both, compared — fit each and reconcile. Skip the prompt when only one data type is present (default to it).

4. Fit. Non-contractual: run scripts/fit_bgnbd.py (BetaGeoFitter MLE → r, alpha, a, b; per-customer expected transactions and probability-alive; then Gamma-Gamma on frequency>0 customers). Validate Gamma-Gamma independence: if corr(frequency, monetary) > 0.3, warn and record potential bias. Contractual: fit Beta-Geometric on renewal-period survival data.

5. Bayesian extension (deep mode). Use pymc_marketing.clv BetaGeoModel and GammaGammaModel; check R-hat < 1.01, ESS > 400; report median CLV with 80% and 95% credible intervals per customer.

6. Predict & score. Run scripts/predict_clv.py for horizons (default 6/12/24 months) with point estimate + interval. Compute CLV = E[transactions] × E[monetary] × margin × discount_factor. If acquisition_costs.json is present, compute CLV:CAC and flag ratio < 1.0; if absent, flag transparently and skip. Tier customers Platinum/Gold/Silver/Bronze by CLV percentile. Build the at-risk list: top-10% CLV × probability-alive dropped below 50% this quarter.

7. Validate. Run scripts/validate_clv.py with a temporal holdout (calibration/holdout split): report MAE, RMSE, and calibration ratio (predicted/actual, target 0.9–1.1).

8. Report. Write outputs (below) and the HTML dashboard.

Output Format

Artifacts:

File	Contents
workspace/analysis/clv_predictions.json	Customer-level CLV with confidence intervals
workspace/analysis/clv_segments.json	CLV tiers (Platinum/Gold/Silver/Bronze)
workspace/analysis/at_risk_customers.json	High-value customers with declining probability-alive
workspace/reports/clv_analysis.html	CLV distributions, cohort curves, model diagnostics

Every readout states: model class, horizon, discount rate, holdout MAE/calibration ratio, and whether CLV:CAC was computed or skipped.

Financial services mode: aggregate CLV across products at the household level; support AUM-basis-point monetary components; compute relationship-level CLV for advisor-intermediated books. Document model inputs for fair-lending review. If CLV outputs feed any customer-facing claim or projected-return copy, route that content through the compliance-review skill before distribution.

Completion status — end every run with one:

• DONE — models fit, holdout passed, artifacts written.
• DONE_WITH_CONCERNS — shipped but note the caveat (e.g., Gamma-Gamma independence violated, weak holdout calibration, CLV:CAC skipped).
• BLOCKED — Hard STOP tripped or required input missing; state the fix.
• NEEDS_CONTEXT — model-class or horizon ambiguity unresolved by the user.

Anti-Patterns

• Emitting CLV numbers from data that failed the integrity gate (future dates, negative amounts, too few repeat buyers).
• Reporting point estimates without confidence/credible intervals.
• Running Gamma-Gamma on one-time purchasers, or ignoring a frequency–monetary correlation above 0.3.
• Presenting CLV:CAC when acquisition-cost data is missing instead of flagging it.
• Using a random (not temporal) holdout — it leaks future information.
• Letting the model narrate CLV or probability-alive instead of running the scripts.
• Defaulting to a contractual or non-contractual model without checking which data the business actually has.