All three contractual deliverables are assembled in the v5.2 package (27 Sep) and mutually consistent. Every quoted figure is produced by one seeded run and checked by a drift gate that reports 0 unmatched on the paper (856 figures), technical supplement (1,119), executive summary (46), quick start (45) and leadership brief (10); the working repository passes 195 tests, ruff and mypy. Files are available on request.
| Deliverable | Version | State |
|---|---|---|
| Peer-reviewed paper CAS_CLV_Paper_v5.2.docx / .pdf (57 pp) · CAS_CLV_Technical_Supplement_v5.2.docx / .pdf (53 pp) |
v5.2 | ✓ complete two documents from the same pinned numbers — paper (conceptual, plain language, 17 figures, glossary App. C) + technical supplement (formulas, full tables, provenance, 13 hand-worked exhibits App. F, 78-term glossary App. G, RFP crosswalk App. H) · v5.2 incorporates Mark Mondello's Appendix F reproduction (R5-1…R5-7, confirmed on Call #9): one toolkit rounding corrected (3+ products 0.9683, mono-line 1.0954), the worked chains print the precision the toolkit divides at, conclusion unchanged · change memo v2.0→v5.2 |
| Executive summary CAS_CLV_Executive_Summary_v5.2.docx / .pdf |
v5.2 | ✓ complete 3 pp — inside the contractual 2–3 |
| Quick start & companion workbook CAS_CLV_Quickstart_v5.2.docx / .pdf (3 pp) · CAS_CLV_Companion_Workbook_v5.2.xlsx |
v5.2 | ✓ complete the workbook reproduces Appendix F in live Excel: 16 sheets, 556 formulas off one Inputs sheet, every check cell zero except the five captioned F.9 rounding differences |
Open-source toolkitcas_clv, MPL 2.0 |
release subset | publication pending curated, independently gated
subset; to be rebuilt from v5.2 and published —
awaiting the casact repository |
| CAS RPM 2027 session proposal + 24-slide deck · submission package |
submission-ready | ✓ built validated against every webform and accessibility limit — awaiting the deadline + submission route |
Five review rounds are answered. Round 1
(AJ Robinson's 18, Mark Mondello's 6) drove v3.0–v3.2; round 2 (Call #6) v3.3; round 3 (Call #7)
v4.0; round 4 (AJ's written review of v5.0 and Mark's first-term expense question, Call #8)
v5.1; round 5 (Mark's Appendix F reproduction, Call #9) v5.2.
Every point either reviewer raised, with its disposition and the file it landed in →
Every point any POG member has raised — in writing, on a call, or in a spreadsheet — with the response, the file the change actually landed in, and its status. One page, kept current, so no reviewer has to hunt across call briefs to find out what happened to their comment. Open the register →
| Reviewer | Raised | Result |
|---|---|---|
| AJ Robinson Allstate |
30 points 18 written on v2.0, 13 Aug · R4-1…R4-12 written on v5.0, 10 Sep |
✓ all closed Drove paper v3.0 — the revision that changed the toolkit and moved reported numbers — and the v5.1 wording round. response memo · register §5c |
| Mark Mondello QBE |
23 points in five waves 3 verbal (Call #3) · 6 written, 17 Aug · 5 spreadsheet cells · R4-13/R4-14, 10 Sep · R5-1…R5-7, 23 Sep |
✓ all closed Drove the tenure-varying loss ratio, the micro-cohort, v3.1, v3.2/L14, the v5.1 first-term expense correction and, through his Appendix F reproduction, v5.2. response memo · register §5d |
| Ronald Kozlowski RTK Services |
1 question, 2 Jul · 93 marked-up comments on v2.0, 7 Sep | ✓ all dispositioned
1 awaiting ratification The expense-timing question
became the split acquisition/renewal expense and clv_net_of_acquisition; the
marked-up review drove v4.0's abstract, §9.6 and the fixed-expense views. RK-46 waits on
POG ratification of the basis names.
response memo |
| Geoff Werner Werner Advisory Aran Paik |
No written review comments received to date | Nothing has been attributed to either member, and nothing invented on their behalf. A pass over paper §§6–7 — the relativity construction, the credibility basis and the admissibility argument — would land on the material with the least independent review so far. Input invited. |
| Tool | What it does |
|---|---|
| CLV scenario planner evergreen · runs in your browser Public link — no account needed → share this one with the POG |
Every lever predict_clv() has, on two modes. Book average values the
fitted 5,000-customer book across all six model families and is an exact factorisation of
predict_clv(), not an approximation — retention shock, both growth hazards,
LAE, the fixed-dollar expense fields, tenure trend on the revenue-neutral or
level-effect basis, horizon to 30 years. Single customer is a discrete term-by-term calculation
on an explicit retention path, so a reviewer can load their own spreadsheet's assumptions
and see exactly where it differs. Presets include the seven named scenarios and the
reviewed worked example, both as written and as corrected.
It checks itself: 100 Python-generated golden vectors are recomputed
in the browser on load, and a test runs the same page's JavaScript under Node against the
Python reference. |
| Review-response register evergreen |
Every point any POG member has raised, its disposition, and the file the change landed in. |
161 register rows across five rounds (164 points, counting Ron Kozlowski's 93 marked-up comments individually): 160 closed, 1 open — RK-46, POG ratification of the revenue-neutral / level-effect basis names, which the manuscript already uses. Four closed rows (c14, c46, MM-A1, RK-39) amend an earlier binding decision and also wait on POG ratification rather than on work. Mark’s Jul-16 request for access without a Hugging Face account closed on 8 Sep, when this hub and the planner became public Spaces.
| # | Date | Phase | Materials | Status |
|---|---|---|---|---|
| — | Jun 8–9 | Kickoff | decisions | ✓ held |
| 1 | Thu Jun 18, 2026 | 1 | Phase 1 framework · brief · slide | ✓ held |
| 2 | Thu Jul 2, 2026 | 2 | brief · framework (upd. §2.1) | ✓ held |
| 3 | Thu Jul 16, 2026 | 3 | brief · worked examples · case studies · POG poll | ✓ held |
| 4 | Thu Aug 6, 2026 moved from Jul 30 · 30 min, hard stop 5:00 PM ET | 4 | brief · cohorts & scenarios · scenario planner · POG poll | ✓ held |
| 5 | Thu Aug 13, 2026 | 5 | brief not used · written review round 1 still received this date (A. Robinson, 18 comments on paper v2.0) · response memo | ⊘ cancelled — rolled into #6 |
| 6 | Fri Aug 21, 2026 full-package review · presented from the RPM deck | 6 | brief · package review · POG poll · session deck · response to M. Mondello · post-call record | ✅ held round 1 accepted · 7 round-2 items opened |
| 7 | Thu Aug 27, 2026 30 min, hard stop 5:00 PM ET | 6 | brief · what the review changed · POG poll · response memo | ✓ held 43 min, all four reviewers present. Reframed the paper for two audiences; asked for an early "so what" illustration, the removal of "price optimization", book valuation as a use case, and one email with draft + links (R3-1 to R3-9) |
| 8 | Thu Sep 10, 2026 30 min, hard stop 5:00 PM ET | 6 | brief · POG poll · deep dive: the one-page case and two expense views · response to R. Kozlowski | ✓ held paper v5.0 + technical supplement · AJ's written review R4-1…R4-12 and Mark's first-term expense question discussed; R4-4 and R4-10 adopted → paper v5.1 (Sep 15) |
| 9 | Thu Sep 24, 2026 | 6 | Mark Mondello's Appendix F reproduction · register §5d | ✓ held R5-1…R5-7 and R4-15 confirmed → paper v5.2 (Sep 27) |
| Date | Decision / input |
|---|---|
| Jun 9 | All 7 kickoff decisions confirmed — hybrid headline (BG/NBD + ML); regulatory CA/TX/NY/NAIC + CO SB21-169 + GDPR brief; both lines, personal-led; weekly written + bi-weekly calls. |
| Jun 18 | Call #1 held — Phase-1 framework, reference walk-through and the 9 framework questions presented (see call materials). |
| Jul 2 | Call #2 — Phase-2 close-out; framework refinement for review: two-variant CLV reporting (unconstrained filing base vs constrained incl. cross-sell + upsell) — POG to confirm terminology. |
| Jul 2 | R. Kozlowski (post-call): incorporate expenses that change over the policy lifetime (acquisition >> renewal)? Accepted — matches brief Q6(a); split acquisition/renewal expense adopted into the framework (§2.2): residual CLV charges renewal expense only (acquisition sunk), inception CLV nets acquisition off; implementation to be shown at the Jul 16 call. |
| Jul 16 | Call #3 — Phase-3 close-out. All six model families live; expense-timing + loss-basis + two-variant work implemented. 5-question POG poll — hybrid guardrails, expense headline, loss basis, CLV labels, comparison-table metrics. Also: worked examples — all six model families on one customer + public-benchmark hand-check (they agree within ±7%). |
| Jul 16 | POG feedback (CLV paper): provide a simplified, visual set of case studies that explain the whole paper at a glance. Built → the whole paper in 6 case studies (1 individual + 5 cohort, per the annuity-analog discipline; inline-SVG visuals, reproducible). |
| Jul 16 | M. Mondello (on call): should losses “differ by how long
you have had them renew” rather than a flat 66% at every renewal?
Accepted & built — CLV is now Σ n_t × margin(t); measured gradient
0.625→0.416 shipped as a smoothed 4.22%/yr credit across all six model families, default
off (exact no-op). Key finding: on the revenue-neutral basis the effect
inverts — the 10+ yr cohort's mean CLV falls 7.4% while younger cohorts rise, so a
naive tenure discount would be backwards. Built →
cohorts & scenarios deep dive |
| Jul 16 | M. Mondello (on call): a very simple case study — “two
people renew, five people renew for ten years” — to show how the model accounts for it
in terms of profit. Built → cas_clv.cohorts.mark_cohort(): seven
named policyholders, every cell hand-checkable; two of seven destroy value yet the cohort
earns +$5,306.98 (+$3,127 as presented on Aug 6; corrected at v5.1,
when the first term took the acquisition ratio in place of the renewal ratio). Scaled to tenure cohorts (mean CLV 4.3× across tenure) and a per-cohort
profit bridge. |
| Jul 16 | Researcher commitment: a web UI where you set loss ratio,
average premium and expenses and read CLV back. Built →
interactive scenario planner (client-side,
all six model families, reconciles to predict_clv() to the cent) + seven named
scenarios and a generated sensitivity appendix. |
| Aug 13 | A. Robinson (written review, 18 comments on paper v2.0): four of them found things that were wrong, not unclear. Cross-sell was being credited in every renewal term rather than once — it is now a one-time event with an annual hazard, and the growth gap falls $693.40 → $376.21 per customer. Credibility was on customer counts against a claim-count standard, which on equal-size quintiles returned an identical 0.9614 in all five rows — a column carrying no information; it is now on claim counts. Built → what the review changed · point-by-point memo |
| Aug 13 | A. Robinson: show the retention-adjusted loss ratio by segment, not by tenure — a by-tenure exhibit cannot support charging less for business expected to persist longer. Built & it is the strongest new result in the paper: the 3+ product segment has the worst first-term loss ratio of three (0.6652) and still indicates a credit (0.9683), on 9.44 expected years against 6.53 for mono-line (paper v5.2; it read 0.9646 on 8.62 against 5.11 at v3.0–v3.2; expected lifetime now comes from the book's persistency table). A single-term exhibit ranks it last of three; the lifetime exhibit ranks it second. Built → Exhibit 11, paper §6.2.1 |
| Aug 13 | A. Robinson: the two variants are named backwards —
“constrained” reads as the regulatorily constrained, filing-safe
variant. Accepted — renamed to the neutral clv_base /
clv_with_growth, with the old names kept as deprecated aliases for one
release so nothing that already runs breaks.
Aug 21 poll Q4 confirms the pair |
| Aug 17 | M. Mondello (written review, 6 comments + a discrete Excel
example): asked for churn, m/r/d and
the BTYD/ML/GBM/GLM abbreviations to be defined at first use; could not follow the
filing-basis inversion; and asked whether the absorbing lapse state double-counts the
non-renewal churn already in Si(t). All six addressed in paper v3.1, and the real-versus-nominal loose end his red cell N17 exposed is resolved in v3.2 (limitation L14)
— exposition only, no figure moved. His spreadsheet reproduces to the cent
($1,759.73 / $1,610.15 / gap $149.58) and independently reproduced the cross-sell
defect, built as it was from the v2.0 text — his per-term cross-sell probabilities
sum to 1.15. Two corrections returned (+$101.23 growth shape, −$199.76 multiplicative
tenure credit → $1,661.20), and one of his red cells found a genuine loose end: day-0
real cash flows discounted at a nominal cost of equity. Built →
response memo + corrected workbook (2026-08-21_CLV_Discrete_Example_reconciled.xlsx) |
| Aug 21 | Full-package review (Call #6). All three contractual
deliverables presented together: paper v3.2 (49 pp at the time; v3.3 is 54), executive summary (3 pp,
inside the contractual 2–3), and the toolkit as a 96-file MPL 2.0 release that
passes its own gate — 130 tests, ruff, mypy and the drift check all clean inside
dist_release/, not merely in the working repo. Presented from the
RPM session deck, whose every figure is generated
from the same seeded payload the manuscript is validated against, so slide and paper
cannot disagree. Asked of the POG → four routing decisions, three of which only CAS
can make: Aug 21 poll |
| Aug 27 | Call #7 (Werner, Kozlowski, Mondello, Robinson; 43 min). Werner: “who are we really writing for?” — two audiences, a business register, and an early “so what” illustration (two customers, one year vs five; Mondello: flat 90% retention is fine for the opener). Werner and Kozlowski: this is not price optimization — drop the term. Kozlowski: could this value a book of business; can the high-value risk be identified ex ante? Robinson: single-line / multi-line for the variants? All: one email with draft and links, not Teams. Register R3-1 … R3-9. Built → paper v4.0: the one-page case, worked |
| Sep 7 | R. Kozlowski (marked-up review of v2.0, 93 comments, Abstract–§5.3). Three changed what the paper says: #24 (“then why write this paper?” — the abstract now states where CLV's economic content is), #95 (book valuation → new §9.6), #169 (the synthetic-data sentence). #112 disagrees with Robinson's #111 on fixed expense → v4.0 carries both a marginal and a fully-allocated view; the default goes to the Sep 10 poll Q2. Global rules on sentence length, jargon and section references applied to the whole paper. All 93 dispositioned (RK-2 … RK-91) → response memo |
| Sep 9 | Researcher → POG: paper v5.0 package — paper (59 pp) and technical supplement (51 pp), one email — both memos, Mondello's reconciled worksheet, planner and hub links. Sent Sep 9 rather than the promised Sep 1–2 so Ron's review could go into one version. No toolkit arithmetic changed; the only new numbers since v3.3 are the §1.1 case ($979.28 / $988.56 today; $3,701.20 / $3,196.62 over five years, gap $504.58) and the fully-allocated example ($782.14 vs $278.88 on the §3.4 renewer; in-force book $13,842,714.22 vs $12,973,898.83) → Sep 10 brief (Corrected at v5.1: the ten-year renewer is $1,093.62 marginal and $590.36 fully allocated.) |
| Sep 10 | Call #8. AJ Robinson's written review of v5.0 (R4-1…R4-12) and Mark Mondello's question on first-term expense (R4-13, R4-14; R4-15 follow-on). R4-4 (the RFP crosswalk moves to Supplement Appendix H) and R4-10 (the toolkit-scope formulation) adopted. Built → paper v5.1 (Sep 15): the first term carries the acquisition ratio in place of the renewal ratio, so the worked example's five-year value is $406.03 and the ten-year renewer $1,093.62 — register §5c |
| Sep 23 | M. Mondello asked for Appendix F as a workbook; building it reproduced every hand-worked chain in live Excel and surfaced seven points (R5-1…R5-7), one of them a toolkit rounding in the segment relativity. |
| Sep 24 | Call #9. R5-1…R5-7 and R4-15 confirmed. Built → paper v5.2 (Sep 27): the segment relativity no longer rounds before dividing (3+ products 0.9684 → 0.9683, mono-line 1.0955 → 1.0954); the companion workbook ships with the package — register §5d |
clv_base / clv_with_growth,
the marginal expense view as default with the fully-allocated view beside it, and rating
decisions as the primary use with book valuation second (§9.6)| Type | Item |
|---|---|
| Friday update | Jun 12, 2026 — first weekly POG update |
| Phase 1 overview | Framework & reference foundations — visual Phase-1 page (framework · how it came from each reference · open questions) |
| Framework | CLV framework · data architecture · dependency decisions |
| Literature | literature review (16-paper corpus) |
| Plan / status | official CAS plan · delivery plan · tracker |
| Sensitivity | sensitivity appendix — eight-dimension grid + seven named scenarios, generated from the seeded book |