Approach summary – Latjor Luk

Ensemble (average of logistic regression, gradient boosting, random forest), seed 42.

Features: prior IPC phase and Phase 3+ %, trend from earlier periods of the same county, county history, population, cereal production/gap, month/season.

Validation: expanding window (train on earlier years only). Ensemble AUC 0.957–0.990 across 2022–2025 vs 0.907–0.986 for a persistence baseline.

Threshold for the write-up: 0.7 (precision 0.99, recall 0.93, specificity 0.89, pooled out-of-time).