Predictability Before Prediction: A Selective-Prediction Framework for Household SKU Choice

Most household product-choice models are evaluated as if a prediction should be issued for every eligible purchase event. This working paper studies a different operational question: can pre-event purchase history identify which household × product events are predictable enough to answer at all? The task is deliberately narrow. Conditional on a household making a purchase within a fixed six-SKU choice set, the target is the SKU selected on that occasion; the study does not model purchase incidence or the composition of a full next basket. Three predictors were developed in a nine-product panel. A market-prior baseline (B0) achieved 37.28% accuracy, a household-history model (B1) 64.53%, and a behavior-fusion model (B2) 65.52% across 17,408 forward-test events. In a separate 26,659-event practical-complexity study, B2 exceeded B1 by 1.2604 percentage points, with positive household- and week-cluster bootstrap intervals, but remained below a predeclared 3.0-point material-gain threshold. The public system therefore retains B1 as the simple predictor. The study then freezes P1, a parameter-free pre-event household-stability score based on history maturity, choice concentration, entropy, switching, and recent concentration, together with three operating cutoffs. In a first new-product batch, the frozen rule achieved 39.16% coverage at 76.18% accuracy, 21.49% at 83.55%, and 11.50% at 88.14%. In a second untouched representative batch containing 40,474 events from 99 evaluable product groups, the corresponding results were 41.08%/77.17%, 22.00%/84.78%, and 11.01%/90.06%. A separate department challenge reached 79.33% accuracy at 40.22% coverage, versus 61.18% for B1 on all events. The evidence supports a simple selective-prediction architecture: use pre-event household stability to decide when to make a conditional SKU-choice prediction and ABSTAIN otherwise. The current evidence grade is CONFIRMED_WITHIN_PANEL_CROSS_PRODUCT; independent retailer replication remains the next test. The result also defines a staged research and application agenda: external replication first, followed by tests of purchase incidence, promotion and price response, decision value, and only then broader commercialization.