← Project hub

CAS CLV Pricing — Bi-Weekly Call Brief

Call: Thursday, June 18, 2026 · 4:30 PM EST (second half) · Call #1 of the bi-weekly series Project: Customer Lifetime Value Pricing (CAS Ratemaking Working Group) Phase: 1 — Literature Review & Framework Design (CAS window Jun 11–24) 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 of all calls / updates / POG inputs: docs/calls/README.md

1. Agenda (proposed, ~30 min)

  1. Progress since kickoff (5 min) — §2
  2. Framework design questions for POG input (15 min) — §4
  3. Schedule alignment + open action items (5 min) — §5, §6
  4. AOB / next call (5 min)

2. Completed since kickoff (2026-06-09)

Project setup & governance - Imported and reconciled the official CAS project plan as the canonical schedule (phase windows + this bi-weekly call series). Repo + QA gate live (pytest / ruff / mypy green; pip install -e . clean). AI-usage log maintained per contract.

Phase 1 — Literature Review (done) - 16-reference corpus retrieved and an annotated literature review written (docs/literature/00_literature_review.md), organized into four streams and mapped to each toolkit module: - Customer-base valuation / marketing CLV — Gupta–Lehmann–Stuart (2004); Verhoef–Donkers (2001) - Insurance survival & household portfolio — Brockett et al. (2008); Guillén et al. (2013) - Multi-state / churn dynamics — Dong, Frees, Huang & Hui (2022) - Probabilistic BTYD & ML — Schmittlein (1987); Fader–Hardie–Lee (2005); Haddadi & Hamidi (2025); Wong et al. (2025); Mahdiyasa et al. (2025); Jena et al.; plus CAS (2004/2014) and Orliński (2026).

Phase 1 — Framework design (draft done) - docs/01_clv_framework.md: formal multi-horizon CLV definition for P&C (CLV_i = Σ S_i(t)·[P − L − E + X]/(1+d)^t), the annuity analog (cohort-level, uncertainty bands — per Orliński), a model taxonomy, and an explicit separation of actuarially-justified vs business-optimization components for filing defensibility (NY §2304).

Phase 2 — Data architecture (ahead of CAS window) - docs/02_data_architecture.md + cas_clv.data: synthetic generator mirroring the Applied Epic model — multi-year, multi-product (auto / home / commercial) customer / policy / claim histories, deterministic under set_seed(42).

Phase 3 — First model (started early) - BGNBD_CLV implemented and tested: a native, scipy-only BG/NBD (Fader–Hardie–Lee 2005), with each policy renewal treated as a "purchase occasion," producing discounted expected renewals × net annual margin. This is the probabilistic base of the POG-selected hybrid headline model. - Engineering decision: implemented BTYD natively rather than depending on the unmaintained lifetimes library — chosen for full auditability (and because its build chain is blocked on the project machine). Documented in docs/03_dependency_decisions.md.

3. In progress / next two weeks

4. Framework questions for POG input

These are genuine design choices where POG steer now saves rework before Phase 3/4 deepen.

  1. CLV horizon basis. Report residual CLV (value of future renewals only, our current default) or total CLV including the current term? Different conventions for filing exhibits.
  2. Reporting grain. Orliński cautions CLV is a relative signal, not a precise dollar value. Do you want filing exhibits at segment/cohort grain with uncertainty bands (our lean), or per-customer point estimates as well?
  3. Discount rate. The proposal uses 8%. Should we anchor on a prescribed/regulatory rate, a risk-free + risk margin, or present discount rate only as a sensitivity (no single headline rate)?
  4. Actuarial vs business-optimization line. Confirm cross-sell / referral value is excluded from rate indications across all target states (CA/TX/NY/CO), or only the strict ones (NY), with the rest allowed to use it. Affects what enters the rate exhibit vs the marketing view.
  5. Headline hybrid model — interpretability vs accuracy. For the "probabilistic form for filing exhibits," how should we weight explainability/defensibility against raw predictive accuracy in the ML-residual layer? (e.g., monotonic constraints, SHAP exhibits, capped residual influence.)
  6. Expense & loss conventions. (a) Single expense ratio (we use 0.25) vs split acquisition/renewal + ULAE/commission breakout? (b) Losses on an ultimate (developed) vs incurred basis — should the synthetic data carry loss development?
  7. Lines of business modeling. Both, personal-led is confirmed. For commercial BOP (different retention dynamics), do you prefer separate models per line or a pooled model with line covariates?
  8. Regulatory depth. For CA Prop 103 and CO SB21-169 disparate-impact testing, what protected- class proxy set and rigor do you expect (which proxies, what thresholds)? This scopes Phase 4.
  9. Worked-example calibration. Should the toolkit be expected to reproduce the proposal's Maria-Chen ~$4,890 5-yr CLV as a validation target, or is that purely illustrative?

5. Schedule alignment (flag)

6. Open action items (carryover from kickoff)

7. Artifacts (available on request / in repo)

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