Interactive CLV scenario planner

CAS Ratemaking Working Group · every lever predict_clv() has, driven from a browser · toolkit cas_clv v1.1 · seed 42

What this is, and which mode you want

Two calculators. They answer different questions, and only one of them can be exact about retention.

Book average

Values the fitted 5,000-customer synthetic book, one model family at a time. A faithful factorisation of predict_clv() — not an approximation. Use it to ask “what does this assumption do to the book?”

Single customer

A discrete term-by-term calculation on an explicit flat retention path. No fitted model, every cell checkable by hand. Use it to ask “does this reproduce my spreadsheet, and where does it differ?”

The page verifies itself: it recomputes a suite of Python-generated golden vectors on load and reports the worst deviation. checking…

Assumptions

Preset

Model family

Tenure basis

Customer lifetime value

Base variant — P − L − E

—

P − L − E only. Defensible in an indication.

With-growth variant — adds cross-sell and upsell

—

Net of acquisition — new business

—

Year-by-year build-up

The same numbers as a table

Year	Expected renewals	Loss multiplier	Effective loss ratio	Discount factor	Base contribution	Growth contribution	Cumulative (with growth)

The footer row is the headline, by construction — base sums to the base card and base + growth sums to the with-growth card. A number you cannot trace to a row is a number you should not trust.

How it computes

CLV = Σ_t=1..H (1+r)^−t × [ (1−e_ren)·A_t·s_P − m_t·B_t·s_L − F·N_t + X_t·N_t ]

A_t, B_t and N_t are the book's premium-weighted, loss-weighted and plain expected renewals for future year t, exported undiscounted so the page applies your own discount rate. s_P and s_L scale them to your premium and loss ratio, F is fixed expense per term, and X_t is the growth weight — cross-sell carrying the one-time first-occurrence weight (1−p)^t−1p and upsell the permanent step-up 1−(1−q)^t.

The tenure multiplier m_t is exact at any horizon and discount rate, not interpolated from a grid. Writing A_t for the loss-weighted mean relative multiplier at full trend, the blend is 1 + trend·(A_t−1) — linear, because the loss weights do not depend on the trend. The revenue-neutral basis then applies one scalar, k = 1/(1+trend·(B−1)), with B the same ratio pooled across the horizon under your discount factors.

Try this. On the revenue-neutral basis, drag the tenure trend from 0 to 1. The two headline figures do not move — but every row of the table does. That is the revenue-neutral property, visible rather than asserted: the curve redistributes expected loss across tenure without changing the book's total. Switch to level-effect and the headline moves, because that basis is a forward-looking assumption about loss improvement rather than a relativity.

Honest limits

Book mode is a book average, not a per-customer prediction. For per-customer CLV, fit a model and call predict_clv().
Retention differs by mode, deliberately. Single-customer mode carries retention exactly. Book mode offers a retention shock, which is the toolkit's own documented book-level approximation — a ratio of flat-retention annuities, because re-estimating per-customer retention would need a refit and a static page cannot refit. A reviewer who wants a precise retention path wants single-customer mode.
Premium and loss ratio scale proportionally. Changing them rescales the fitted book rather than refitting it, so the renewal pattern is held fixed.
The 30-year horizon is an extrapolation past about year 10. The synthetic book carries roughly ten years of history; beyond that the renewal pattern is the model's extension of its own fit, not something the data evidences.
The additive tenure convention is not the toolkit's. It is offered because a reviewer's spreadsheet may read the credit as loss-ratio points per year, and seeing the difference is the point. The toolkit compounds.
No price elasticity anywhere. There is no rate-change or conversion lever and there will not be one — a test asserts that the survival covariates cannot gain one. That invariant is what keeps the filing-side exhibits free of demand-based content.
Synthetic book. Coefficients come from 5,000 seeded synthetic customers, not a real carrier.

Payload and golden vectors generated by python scripts/build_clv_planner.py (set_seed(42)), which refuses to write unless the Python reference in cas_clv.planner_reference reconciles to predict_clv() across a sweep of lever combinations and reproduces the reviewed worked example both as written and as corrected. Related: review-response register · cohort deep dive.