CAS CLV Pricing — Bi-Weekly Call Brief

Call: Thursday, July 16, 2026 · 4:30 PM EST · Call #3 of the bi-weekly series Project: Customer Lifetime Value Pricing (CAS Ratemaking Working Group) Phase: 3 — CLV Model Development & Comparison (CAS window Jul 1–14) — closing complete Researcher: Pramod Misra (5G Vector partner: Neha Tiwari) CAS staff / coordinators: Elizabeth Smith (), Heather Davis () Project Oversight Group (reviewers): Mark Mondello (QBE), AJ Robinson (Allstate), AJ Paik, Geoff Werner (Werner Advisory), Ronald Kozlowski (RTK Services) Master log: docs/calls/README.md

1. Agenda (proposed, ~30 min)

Phase 3 close-out: all six model families live (10 min) — §2
Expense-timing implementation demo — follow-through on Ron's question (5 min) — §3
Research corpus + public-dataset repository (3 min) — §4
Questions for POG: hybrid interpretability + comparison metrics (7 min) — §5
Action items / next call Jul 30 (5 min)

2. Phase 3 — ✅ complete inside the CAS window

All model families from the proposal are implemented, tested, and expose one common contract — fit() / predict_clv() / sensitivity_analysis() — with every model reporting the two CLV variants (§2.1) and the residual/inception expense conventions (§2.2):

Family	Toolkit class	Notes
Contractual probabilistic	`BGNBD_CLV`, `ParetoNBD_CLV`	Native scipy closed forms (Fader–Hardie–Lee 2005; Schmittlein 1987); renewal = purchase occasion
Survival	`SurvivalCLV` (Cox PH · Weibull AFT)	Native; conditional residual survival; lapsed = 0
Multi-state	`MarkovCLV`	Product-portfolio states + absorbing lapse (Dong–Frees 2022 template); multi-product states demonstrably stickier
Latent-state	`MarkovCLV(n_latent_states=k)`	Gaussian-HMM engagement states over yearly product/tier observations
Predictive ML	`EnsembleCLV(mode="ensemble")`	XGBoost + LightGBM on customer-disjoint temporal splits
Hybrid (headline)	`EnsembleCLV(mode="hybrid")`	BG/NBD probabilistic base + GBM residual correction — the form POG selected at kickoff; filing exhibits keep the probabilistic base

Retention engines agree across families (BTYD cross-validation rank corr > 0.8; survival and Markov consistent with the book's empirical retention).
Everything deterministic under set_seed(42); QA gate green (60+ tests / ruff / mypy).
Synthetic book upgraded to generator v2 with a CI-enforced realism gate: every headline statistic (premiums, frequencies, loss ratios ~53%, retention 0.924, margin ~22%) is asserted against public-benchmark bands on every build, plus a parameter-recovery test.

3. Expense timing — implemented as promised (R. Kozlowski, Jul 2)

Split acquisition_expense_ratio (0.40) / renewal_expense_ratio (0.20), both sensitivity dimensions; residual CLV charges renewal expense only (acquisition sunk); inception CLV nets acquisition off. On the synthetic book: flat-0.25 CLV $2,369 → split residual $2,943 → inception $1,341 per customer — the recoupment story, quantified.
Live demo available: notebook 01_data_exploration.ipynb exhibit + the Phase-4 expense-recovery schedule (breakeven year) already prototyped.

4. Research foundation expanded

Reference corpus 16 → 30 items (adds: price-optimization/elasticity stream incl. the NAIC 2015 white paper; CO Reg 10-1-1 / NAIC AI Model Bulletin / NYDFS CL-7 primary sources; Fader–Hardie technical notes; ML-pricing benchmarks).
Public-dataset repository (research/data/): 7 datasets (French MTPL2, Wisconsin LGPIF panel — the Dong–Frees fund, Spanish multi-product personal lines with retention flags, CAS Schedule P auto triangles, Australian + Swedish motor) + a fully-cited industry-benchmark table backing every calibration band. Enables real-data validation variations in Phase 5.

5. Questions for POG input

Hybrid interpretability vs accuracy (carryover): the toolkit now ships a feature_importances() exhibit for the ML layer — on the reference book it is dominated by exposure and renewal history (tenure ≈ 0.40, renewal frequency ≈ 0.25, premium ≈ 0.14), i.e. the residual learner sharpens retention signal rather than importing new rating variables. Is this exhibit + the probabilistic-base-for-filings design sufficient, or do you also want monotonic constraints / capped residual influence / SHAP?
Model-comparison table metrics: which columns do you want in the paper's headline table — holdout MAE on discounted renewals, rank correlation vs realized value, calibration of 1-yr renewal, lift/Gini by decile, runtime?
Expense defaults: acquisition 0.40 / renewal 0.20 (blend ≈ industry 25–27% expense ratios) — acceptable as published defaults? [POG-FEEDBACK]
Loss basis (brief Q6b, now decision-ready): the synthetic claims now carry BOTH bases — ultimate (incurred_loss, default) and as-of-date reported (reported_loss, via stylized Schedule P development patterns) — and every model accepts loss_basis="ultimate"|"reported". Which convention should the published exhibits lead with? (Our recommendation: ultimate for pricing exhibits, with a reported-basis sensitivity in the appendix.)
[POG-FEEDBACK: carry forward anything still open from the Jul 2 call — two-variant labels, discount-rate convention, residual-vs-total horizon.]

6. Open action items

CAS: ScholarOne + github.com/casact provisioning [CONFIRM status]; RPM date [CONFIRM].
Researcher: Phase 4 (ratemaking + regulatory) begins Jul 15 per plan.

7. Artifacts

src/cas_clv/ (all 6 families) · tests/ (QA green) · notebooks/01_data_exploration.ipynb · research/papers/ (30) · research/data/ (7 datasets + benchmarks) · framework doc §2.1/§2.2.

This research project has been funded by the Casualty Actuarial Society.