Abstract / Summary
Abstract Uneven coverage is a basic constraint in food-price surveillance. Analysts must rank country-month signals even when the number and composition of market quotes vary across place and time. This study examines how that ranking depends on the price measure, the event state that defines eligibility, and the rule used to order cases. The primary index aggregates trimmed monthly changes calculated within WFP market-commodity-unit-currency series. Past-only thresholds then define one-month-ahead shock onset, high-risk status, and extreme-shock status. Split-half reconstructions assess the stability of the index and its labels. Matched rolling-origin evaluations for 2019–2021 and 2022-April 2026 compare nested logit and LightGBM selection with pooled-event training, an inherited fixed architecture, and transparent price-history rules. Official Food CPI series and public IPC country-period estimates provide independent criteria. Reliability improves with quote coverage, while the leading rule varies by task. For recent high-risk status, nested target-specific selection reaches a PR-AUC of 0.246, compared with 0.222 for pooled-event training, 0.198 for the inherited fixed architecture, and 0.229 for threshold distance, the strictest transparent rule. Its direct advantage over threshold distance is 0.017, with a selection-aware 95% interval of [-0.029, 0.048]. Pooled-event training leads recent onset, threshold distance leads recent extreme shocks, and price-history rules lead the historical tasks. The retail-price stress measure also ranks later five-percentage-point deterioration in the IPC Phase 3 + population share above sample prevalence. Surveillance performance therefore turns on choices made before model fitting: the price measure, the eligible state, and the review workload. The resulting scores support expert assessment by ordering cases within that workload.