hisprwanda/sarimax-xgboost-comoros-malaria-prediction-model

Ensemble SARIMAX–XGBoost Model for Malaria Forecasting in Comoros # Comoros Malaria Climate-Health Forecasting Probabilistic weekly malaria forecasting for the **Ngadjizi region, Comoros** (7 districts), using climate drivers — rainfall, temperature, and humidity. Built on the CHAP (Climate Health Analytics Platform) model specification. Models are fitted independently per district and produce calibrated probabilistic forecasts as sample distributions. --- ## The Problem Malaria transmission in Comoros is strongly seasonal, driven by the Indian Ocean climate. The **hot season is the wet season**: high temperatures heat the ocean surface, driving evaporation and rainfall. This means temperature and rainfall are positively co-seasonal — a warm week predicts a wet week 1–3 weeks later, and both predict a mosquito breeding surge 1–4 weeks after that. Early, calibrated forecasts let district health teams pre-position rapid diagnostic tests, antimalarials, and bed nets *before* a transmission peak — not after it. --- ## What This Project Does An 8-experiment research series that systematically builds and improves a probabilistic malaria forecasting system: | Step | What was done | |---|---| | Baseline benchmark | 7 models evaluated across 78 weeks train / 26 weeks test | | Exp 01 | Per-district climate–case cross-correlation & lag analysis | | Exp 02 | SARIMAX auto-ARIMA order selection vs fixed (1,0,1) | | Exp 03 | District-specific informed feature selection (\|r\| > 0.10 threshold) | | Exp 04 | Prophet hyperparameter tuning (40-config grid search) | | Exp 05 | XGBoost calibration fix — residual bootstrap → quantile regression | | Exp 06 | Rebuilt S+P+X ensemble with all tuned components | | Exp 07 | **S+X ensemble discovered as champion** — Prophet removed from pool | | Exp 08 | **Operational robustness** — CRPS degradation under climate forecast noise | --- ## Champion Model — Ensemble S+X | Metric | Value | |---|---| | **CRPS** | **25.91** | | **RMSE** | **48.84 cases/week** | | **80% PI Coverage** | **76.9%** | | **95 …