Completed since kickoff (Jun 9)
SETUPOfficial CAS plan adopted as canonical schedule;
repo + QA gate live (pytest/ruff/mypy green); AI-usage log maintained.
PHASE 1 ✓Literature review done.
- 16-paper corpus retrieved + annotated synthesis, mapped to every toolkit module
- 4 streams: customer-base valuation · insurance survival · multi-state churn · BTYD/ML
PHASE 1 ✓Framework drafted.
- Formal multi-horizon CLV definition: CLV = Σ S(t)·[P−L−E+X]/(1+d)t
- Annuity analog (cohort + uncertainty bands); model taxonomy
- Actuarially-justified vs business-optimization split (filing-defensible)
AHEADPhase 2 data + Phase 3 first model.
- Synthetic Applied-Epic-mirroring generator (auto/home/commercial)
- BG/NBD CLV model built & tested — the hybrid headline's probabilistic base
- Native, auditable implementation (no fragile dependencies)
NEXT 2 WKSPareto/NBD + Cox/AFT survival · first worked
notebook · calibrate to the ~$4,890 worked example.
Framework questions for POG input
- Horizon basis — residual (future renewals) vs total CLV incl. current term?
- Reporting grain — cohort + uncertainty bands vs per-customer point estimates?
- Discount rate — prescribed/regulatory, risk-free+margin, or sensitivity-only? (8% in proposal)
- Cross-sell value — excluded from rate indications in all target states, or only NY?
- Hybrid model — how to weight interpretability/defensibility vs accuracy in the ML layer?
- Expense & loss — single expense ratio vs split; losses ultimate vs incurred?
- Lines of business — separate models per line vs pooled with line covariates?
- Regulatory depth — proxy set & rigor for CA Prop 103 / CO SB21-169 disparate-impact testing?
- Calibration — reproduce the ~$4,890 5-yr worked example as a validation target, or illustrative?