SENTINARA

Model Lab

Every forecast and warning on SENTINARA comes from models tested in the open: backtests, calibration, a real-time forecast archive, climate tests and multi-year cycle analysis.

Key facts

  • Indonesia one-month-ahead backtest over 26 months: MAPE 12.6%, MASE 0.25 (below 1 beats a seasonal naive forecast), 80% interval coverage 81%.
  • Pooled calibration across 54 countries: 50% ranges held 58%, 80% ranges 80%, 90% ranges 89% and 95% ranges 94% of the time.
  • Best single model by country: harmonic regression with AR terms 26, damped Holt-Winters 9, seasonal naive 3.
  • Real-time archive started 2026-10-30: 55 forecasts stored; the first forecast for each month is kept so later data cannot rewrite it.
  • Climate test: rainfall leads Indonesian dengue by about 1 month(s) (r = 0.383), temperature by 3 (r = 0.439) and El Niño (ONI) by 5 (r = 0.292). Adding climate did not improve the backtest (MAPE 12.8% with, 12.6% without), so it is not used in the national forecast. Current ONI 2.16 (strong el nino).
  • Multi-year cycles: Indonesia's strongest cycle is about 10.6 years but it is not statistically significant (p = 0.128). Significant cycles were found for Australia (5.2 years), Guyana (11.6 years), New Caledonia (4.8 years).

Data updated 2026-10-31. Sources: WHO Global Dengue Surveillance; OpenDengue V1.3 (Clarke et al., 2025), doi:10.6084/m9.figshare.24259573; Dinas Kesehatan Kota Semarang; World Bank Open Data; NASA POWER.