Five quick decisions for the group — each one is a choice on work you already steered.
The green banner shows what you asked → what we built; then pick the option
you want carried into the paper and the toolkit.
★ = researcher's recommendation.
Q1
Hybrid model — how much interpretability armor?
✓You asked (kickoff): make the
Hybrid the headline model. ✓We built it
(BG/NBD base + ML residual) and added a feature_importances() exhibit —
the ML layer is dominated by tenure ≈0.40 and renewal frequency ≈0.25,
i.e. it sharpens retention, it does not smuggle in new rating variables.
For filing-grade defensibility, how far should the guardrails go?
A
Feature-importances exhibit +
keep the probabilistic base for filings — sufficient★ rec
B
Add SHAP explanations per prediction
C
Add monotonic constraints / cap the residual's influence
D
All of the above — maximum regulatory armor
Q2
Expense timing — which view do we publish?R. KOZLOWSKI
✓You asked (Jul 2): should expenses vary
over the lifetime? ✓We built it — split
acquisition 0.40 / renewal 0.20, residual CLV charges renewal expense
only (acquisition is sunk). On the book: $2,369 flat → $2,943 residual → $1,341 inception.
Which should be the published headline CLV?
A
Residual CLV (renewal expense only)
— the retention-value story★ rec
B
Inception / new-business CLV (nets acquisition off)
C
Show both side-by-side, flat ratio as fallback
D
Revisit the 0.40 / 0.20 ratios first
Q3
Loss basis — ultimate or reported?
✓You flagged ultimate vs reported losses
(brief Q6b). ✓We built both — every model takes
loss_basis = "ultimate" | "reported", reported via stylized Schedule P
development patterns. This is a pick, not a rework.
Which convention leads the published exhibits?
A
Ultimate for pricing exhibits +
reported-basis sensitivity in the appendix★ rec
B
Reported / as-of-date leads
C
Present both as co-equal
D
Defer — show me the numbers under each first
Q4
Two-variant CLV — do the labels work?
✓Design agreed (Jul 2): report a
filing-defensible renewal base and a planning value that adds cross-sell + upsell.
✓Shipped as clv_unconstrained /
clv_constrained on every predict_clv() — you asked us to confirm the naming.
Which labels go into the paper?
A
Keep "unconstrained" / "constrained"
B
"Filing CLV" / "Planning CLV"★ rec
C
"Base CLV" / "Total CLV"
D
"Renewal-only" / "With cross-sell"
Q5
Model-comparison table — which metrics matter most to you?
✓Built: all six model families
(BG/NBD · Pareto/NBD · Cox/AFT survival · Markov · HMM · Ensemble/Hybrid) run on
one common predict_clv() contract, so they compare apples-to-apples.
Now you choose the columns for the paper's headline comparison table.
Which metric set do you want reviewers to see?
A
Core: holdout MAE on discounted
renewals + rank correlation vs realized value★ rec
B
Add 1-year renewal calibration
C
Add lift / Gini by decile
D
Add runtime / operational cost