Methodology

Where the numbers come from, what was done to them, and what they cannot tell you.

Where each number comes from

Pakistan is served from a locally rasterised CMIP6 0.25° grid — 1,512 extracted fields — so a point value is a real grid cell. Uzbekistan, Australia and New Zealand have no local raster yet. For those, the magnitude is the published CCKP national value (1,344 values fetched from the aggregate API and stored in data/cckp-national.json), and the variation within the country is interpolated from elevation and latitude. Those readouts are labelled interpolated, never presented as grid cells. Run pnpm grid:extract to replace the interpolation with real rasters.

Read this before quoting an individual model

The 30-member ensemble is published only as a national spatial aggregate. Gridded fields exist for the ensemble percentiles (median, 10th, 90th), not per model. So on this platform the model-spread envelope at a point is local, but the per-model list is a country-wide average — not a grid-cell value at the selected coordinate. Every panel that shows individual models says so. This is the limitation a reviewer should probe first.

Data sources

CMIP6 (downscaled, 0.25°)

projection

Answers: How could the climate of a place change over multi-decade windows under a given emissions pathway?

Provider: World Bank Climate Change Knowledge Portal
Resolution: 0.25° × 0.25° (~25 km)
Coverage: 1950–2100

ERA5 reanalysis (0.25°)

observation

Answers: What has the climate of a place actually been?

Provider: ECMWF, via World Bank CCKP
Resolution: 0.25° × 0.25°
Coverage: 1950–present

Open-Meteo forecast

forecast

Answers: What will the weather do over the next few days?

Provider: Open-Meteo
Resolution: ~1–11 km, model dependent
Coverage: now + 16 days

These three are never combined. A short-range forecast, an observational reanalysis and a multi-decadal projection answer different questions with different methods and different error characteristics. They are kept in separate stores, separate endpoints and separate parts of the interface for that reason.

Processing pipeline

Raw CMIP6 output is not usable directly: it arrives as global NetCDF rasters, one file per variable, model, scenario, product, aggregation, percentile and period. The pipeline reduces those to the smallest artefact that can still answer a point query.

1
Source
Global 0.25° NetCDF fields, ~8 MB each, from the World Bank CCKP open data bucket on S3.
2
Subset
Clip to 60.5–78.0 °E, 23.5–37.25 °N — a 71 × 56 lattice of 3,976 cells covering Pakistan including Azad Jammu & Kashmir and Gilgit-Baltistan.
3
Reduce
Flatten to a stacked array indexed layer-major and then row-major from the south-west corner, with the grid geometry stored once rather than per field. Annual fields have one layer, seasonal four and monthly twelve. A national annual field compresses to roughly 11 KB.
4
Index
Precompute which grid cells fall inside each of the 7 provinces and 126 districts, using point-in-polygon against the unsimplified boundaries so no border cell is lost to simplification.
5
Serve
Point queries read the lattice directly. Anything not rasterised locally falls back to the CCKP aggregate API, and the response says which source answered.

Deliberately, the database is not a scientific data warehouse. Putting every raw NetCDF value into Postgres would mean hundreds of millions of rows to answer questions that a 25 KB array answers in microseconds. The pipeline computes application-ready products; the store serves them.

Coverage in this deployment

Rasterised fields

1,512

Grid payload

17.99 MB

Grid cells

3,976

Locally rasterised indicators: cdd, cdd65, hd35, hd40, hi35, pr, r95ptot, rx1day, rx5day, sd, tas, tasmax, tasmin, tnn, tr23, txx. Anything outside that list is still fully queryable — it is served from the upstream aggregate API as a national average, and every response and panel says so explicitly rather than presenting a country-wide mean as a local value.

Grid last built Wed, 02 Sep 2026 15:58:37 GMT.

Analytical choices

Ensemble median as the default

The 10th and 90th percentiles are percentiles across models, not confidence intervals. An individual model carries only a mean, because a cross-model percentile of a single model is not a meaningful quantity — the API rejects that combination rather than silently returning something.

Anomalies against 1995–2014

Models agree far better on change than on absolute values. The baseline is the one CCKP publishes every anomaly against, so anomalies here are the upstream product rather than a subtraction done locally, and no bias is introduced by re-differencing.

Equal-weighted area aggregation

Province and district means weight every grid cell equally. A 0.25° cell varies about 15% in area between Karachi and Gilgit; that is small next to the model spread already shown, and equal weighting keeps values reproducible against CCKP’s own aggregates.

Colour scales clipped to the 2nd–98th percentile

A handful of extreme Karakoram cells would otherwise flatten contrast across the plains where most people live. Diverging scales are forced symmetric about zero so that equal changes of opposite sign are equally saturated.

Units taken from the files, not from the variable names

Several CCKP variable codes do not mean what they look like.sdis the ETCCDI “summer days” count, not snow depth;r95ptotis a percentage share of wet-day rainfall, not a millimetre total;cdd65is published in Fahrenheit degree-days; and the precipitation anomaly is in millimetres rather than percent. Every unit shown on this platform is taken from theunitsattribute of the source file and pinned by a test.

Model disagreement shown, not hidden

Where CCKP publishes an agreement classification, cells in which models conflict on the sign of change are drawn faded. The median is still the median; its direction is simply not robust there.

Limitations

25 km cells cannot resolve a city

A grid cell spans roughly 25 km. Urban heat islands, valley inversions and coastal breezes all operate below that scale. Karachi’s cell describes the region, not the neighbourhood.

Complex terrain is where downscaling is weakest

In the Hindu Kush–Karakoram–Himalaya, elevation changes by kilometres within a single cell. Snow, precipitation phase and glacier mass balance are all sensitive to exactly that, and are correspondingly less reliable in the north than on the plains.

The South Asian monsoon is a known model weakness

CMIP6 models disagree more on monsoon precipitation than on almost any other regional signal, and aerosol forcing — which differs between SSPs by construction — has a first-order effect on it. Precipitation projections for Pakistan deserve markedly more caution than temperature ones.

Twenty-year means say nothing about a given year

Every value here is a climatology. A projection of +2 °C for 2040–2059 does not mean 2047 will be 2 °C warmer than 2002.

Time series are national only

The archive publishes continuous annual traces as spatial aggregates. Any chart running 1950–2100 on this platform describes Pakistan as a whole, and is labelled accordingly.

No impact modelling

This platform reports climate variables. It has no hydrological, crop, health or economic model, and makes no claim about floods, yields or mortality.

Scope

Scenarios: 5 SSP pathways; SSP1-1.9, SSP1-2.6, SSP2-4.5, SSP3-7.0, SSP5-8.5
Models: 30 GCM realisations + ensemble; Bias-corrected and statistically downscaled by the World Bank.
Periods: 5 windows; 1995–2014, 2020–2039, 2040–2059, 2060–2079, 2080–2099
Geography: 7 provinces, 126 districts; Boundaries from geoBoundaries (gbOpen, public domain), simplified for web delivery.

Citation

World Bank Group. Climate Change Knowledge Portal — CMIP6 bias-corrected downscaled projections, 0.25°. Available at climateknowledgeportal.worldbank.org.
Eyring, V. et al. (2016). Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geoscientific Model Development 9, 1937–1958.
O’Neill, B. C. et al. (2016). The Scenario Model Intercomparison Project (ScenarioMIP) for CMIP6. Geoscientific Model Development 9, 3461–3482.
IPCC (2021). Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report.
Runfola, D. et al. (2020). geoBoundaries: A global database of political administrative boundaries. PLoS ONE 15(4).

Suggested citation

Earth Scan Systems (2026). Climate Intelligence Platform: CMIP6 downscaled projections for Pakistan, Uzbekistan, Australia and New Zealand. Data: World Bank Climate Change Knowledge Portal.

@misc{ess_climate_platform_2026,
  author       = {{Earth Scan Systems}},
  title        = {Climate Intelligence Platform: CMIP6 downscaled
                  projections for Pakistan, Uzbekistan, Australia
                  and New Zealand},
  year         = {2026},
  note         = {Data: World Bank Climate Change Knowledge Portal,
                  CMIP6 bias-corrected downscaled projections (0.25{\deg}).
                  Baseline 1995--2014.},
  howpublished = {\url{https://climateknowledgeportal.worldbank.org/}}
}