Skip to content

Climate scenario outlook

Under testing

The scenario outlook is under active testing. It has been checked only on synthetic climate-model data with a known answer. It could not yet be run on the real CMIP6 files in development. Projections are "what if" estimates, not predictions of actual future cases.

DiseaseModel.project_v2() estimates how cases could change under future climate, for each CMIP6 climate model and emissions pathway (SSP), from next season out to 2050 (or 2100).


How the outlook is built

  1. Bias correction. Each climate model is shifted so its baseline-period monthly averages match observed climate (rainfall scaled, other variables shifted). Its projected change is kept.
  2. Next 12 months: the v2 forecast, driven by each climate model's corrected climate.
  3. Hand-over (months 13–24): a gradual switch as recent cases stop being informative.
  4. Long term: a model that links cases to climate alone, refitted many times on resampled years so the ranges include statistical uncertainty.
  5. Pooling: results from all climate models are combined, so ranges also include climate-model spread.
  6. Checks reported alongside: a sensitivity run using a different way of learning climate effects, a backtest on held-out past years, and a comparison with tree-based models.

Main options

Option Default Meaning
end_year 2050 Last year of the outlook
ssps all available e.g. ["ssp126", "ssp245", "ssp585"]
response "seasonal" Learn climate effects from the seasonal cycle ("anomaly": only from year-to-year deviations)
lag_selection "ensemble" "ensemble": average all near-equally-good climate-lag structures. "v1": run ClimAID v1's lag search (in v1_mode) on data before the forecast start. "distributed": smooth lag curves over 0–6 months (ENSO 0–12)
v1_mode from tuning v1 mode for lag_selection="v1": "fast", "balanced", "deep"
extra_districts None Other districts' data: learn one shared climate response (recommended if available)
temperature_curve None e.g. "aedes_aegypti_mordecai2017": constrain the temperature response to a thermal-suitability curve
population_projection None Table of projected population; adds population-scaled results
comparison_models random forest, gradient boosting Tree models shown as a comparison
tuning "balanced" Tuning of the near-term and comparison models
exclude_period "2020" COVID-19 period left out of the long-term fit and replaced for the near-term forecast

Choosing the lag structure

Tested on synthetic cases with a known warming effect (change by the 2050s):

Case (truth) Averaged (default) v1 lags Distributed lags
Seasonal benchmark, SSP5-8.5 (+60%) +27% +48% +58%, but range up to +200%
Six-district world, single district (+27%, 3 seeds) +13 / +16 / +30% — +14 / +41 / +54%, wide ranges
Six-district world, pooled (+27%, 3 seeds) +25 / +24 / +30% — +41 / +33 / +34% (overestimates)

No single choice was best everywhere. The averaged default was the most reliable when districts are pooled; distributed lags estimated the total effect of warming well in one case but were unstable in others. The fitted distributed-lag curves recover the total effect better than the timing, so do not read the curve shape as "when the effect happens".

Known limitations (read before using)

Single-district projections can understate warming effects

Temperature, rainfall and humidity often share one seasonal cycle, so a single district's data cannot fully separate their effects. On synthetic data where warming truly raised cases by 60% by the 2050s, the default single-district projection gave +27% (the truth was inside its stated range). Pooling six districts with extra_districts recovered the truth within 3 percentage points. Use extra_districts, lag_selection="v1" or a justified temperature_curve where you can.

  • The climate–disease relationship is assumed to stay the same in future.
  • Population (unless scaled), immunity, vector control, testing and reporting are held at historical levels.
  • Climate models represent El Niño poorly; its effects mainly add noise far ahead.
  • Where future climate goes beyond the range seen in training, the response is extrapolated (the report says how often this happens).
  • The default thermal-curve preset values (17.8 / 29.1 / 34.6 °C) are taken from Mordecai et al. (2017).