Skip to main content
v2026.11,986 entries · CC-BY 4.0

Downscaling: Dynamical vs Statistical Methods Explained

What downscaling is, how dynamical and statistical methods differ, how CORDEX organises regional runs, and what to check before reusing downscaled data.

Written and maintained by CASRAI Editorial Board

Published

On this page

Downscaling is a method that derives local- to regional-scale climate information from larger-scale models or data analyses. The IPCC names two main routes: dynamical downscaling, which runs a regional climate model (or a variable-resolution or high-resolution global model) driven by coarser output, and empirical/statistical downscaling, which builds statistical relationships between large-scale atmospheric variables and local climate variables from observations. In every case the quality of the driving model remains an important limit on the quality of the result.

Short answer

If a global climate model is too coarse for your question, downscaling is how regional and local information is produced from it. Dynamical downscaling adds physics at finer grid spacing; statistical downscaling adds observed local relationships. Neither one repairs errors in the large-scale circulation of the driving model, and a finer grid does not automatically mean better information. The rest of this guide explains the definition, how the community-wide CORDEX programme organises dynamical downscaling, how the two methods differ in what they need, and what to check before reusing a downscaled dataset.

Definition and scope

The IPCC Sixth Assessment Report (Working Group I, Annex VII glossary) defines downscaling as a method that derives local- to regional-scale information from larger-scale models or data analyses. The same entry separates two main methods and treats “dynamical downscaling” and “statistical downscaling” as pointers to that single entry.

  • Dynamical downscaling uses the output of regional climate models, global models with variable spatial resolution, or high-resolution global models.
  • Empirical/statistical downscaling is based on observations and develops statistical relationships that link large-scale atmospheric variables with local or regional climate variables.
  • Combined approaches are possible. The glossary gives the example of applying empirical/statistical downscaling to the output of a regional climate model, which is itself a dynamical downscaling of a global climate model.

Downscaling is a way of working with climate model output and analyses, not a separate climate model of its own. Related inputs include reanalyses, which the IPCC describes as past observations processed with a fixed forecasting model and data assimilation, and which “still suffer from changing coverage and biases in the observing systems”. Climate projections, as opposed to predictions, depend on an emission or concentration scenario, so any downscaled projection inherits that scenario dependence from the global model that drives it. For where such global and regional climate data are distributed, see the Copernicus Climate Data Store guide.

How downscaling is done in practice: the CORDEX example

The most widely used coordinated framework for dynamical downscaling is CORDEX, the Coordinated Regional Downscaling Experiment, which sits under the World Climate Research Programme (WCRP). Its published experiment design for downscaling CMIP6 global model output (28 April 2025, DOI 10.5281/zenodo.15268192) states that the aim is to improve downscaling techniques and their use in understanding and assessing regional climate change, and to provide regional information for vulnerability, impacts and adaptation studies. The CMIP5-based phase of CORDEX has been running since 2010.

What a CORDEX-CMIP6 dataset specifies

Reading a downscaled dataset’s metadata is easier when you know the design it follows. The experiment design document sets out the following (the data table below lists them with sources):

  • Domains. 14 standard continental-scale domains, including the Arctic and Antarctic.
  • Grid spacing. The primary targets are 25 km and 12.5 km, with one common spacing recommended per domain. Spacing finer than 12.5 km is allowed by agreement, alongside complementary coarser runs.
  • Minimum model. A regional atmosphere-land surface model.
  • Evaluation run first. Every modelling group first runs a simulation driven by the ERA5 reanalysis. The ERA5 period is 1979-2024 with a one-year spin-up starting in 1979, and ERA5.1 is recommended for 2000-2006 because of a cold bias in the lower stratosphere in ERA5.
  • Scenario runs. Historical CMIP6 boundary conditions for 1950-2014, then greenhouse-gas forcing following SSP3-7.0 from 2015 onward.
  • Which global models to downscale. The document states there is no commonly accepted methodology for selecting which global climate models to downscale and leaves the choice to regional communities.
  • Distribution. Output is published on the Earth System Grid Federation (ESGF) to the CORDEX-CMIP6 archive specification.
  • Spectral nudging. Whether to use it is left to the regional-model group.

These details matter for reuse: a dataset produced under this design has a known evaluation run, known driving-model boundary conditions and a known archive format, so another team can check what was done rather than guess. Most CORDEX output is delivered as gridded files; see the NetCDF guide and the Climate Data Operators guide for handling them.

Dynamical versus statistical versus combined

Gutowski and colleagues (2016, Geoscientific Model Development 9, 4087-4095) describe regional climate downscaling as including both dynamical and statistical approaches, and note practical differences in what each needs. Dynamical downscaling needs six-hourly global model fields as forcing, over at least 1950-2100 in the CORDEX design. Statistical downscaling “has more flexibility with input data” and, in the CORDEX-CORE plans, may go directly to observation sites. The table below sets the two approaches side by side using only statements from the IPCC glossary and that paper.

Question Dynamical downscaling Empirical/statistical downscaling
What it is built from Output of regional climate models, variable-resolution global models or high-resolution global models (IPCC) Observations, used to derive statistical relationships between large-scale and local variables (IPCC)
Input data it needs Six-hourly global model forcing fields, 1950-2100 minimum in the CORDEX design (Gutowski et al. 2016) More flexible about input data (Gutowski et al. 2016)
Where results can be produced On a regional model grid, for example 25 km or 12.5 km in CORDEX-CMIP6 (CORDEX design) Can go directly to observation sites (Gutowski et al. 2016, CORDEX-CORE plans)
Main stated limit Quality of the driving model remains an important limitation (IPCC) Quality of the driving model remains an important limitation (IPCC)
Can it be combined with the other? Yes: its output can be further downscaled statistically (IPCC) Yes: it can be applied to regional climate model output (IPCC)

The shared limit in the fourth row is the practical point. Both methods take a coarser source as given, so the choice between them is about the question and the available inputs, not about escaping the driving model.

Where it matters for research data and model-output reuse

For a researcher or data steward, downscaled output is a derived dataset with a long provenance chain: observations or a reanalysis, a global model, a scenario, a downscaling method, and sometimes bias-handling steps. The sources above support a short checklist for deciding whether such a dataset is fit for a specific use.

  1. State the user problem first. The VALUE framework (Maraun et al. 2015, Earth’s Future 3, 1-14) is a user-focused approach to validating downscaling: it starts from a specific user problem and uses a validation tree to guide the choice of indices and measures.
  2. Separate method skill from inherited error. VALUE’s experiments are designed to isolate where problems arise: the isolated skill of the downscaling, statistical versus dynamical methods, different spatial scales, failure to represent regional climate change, and overall representation including errors inherited from the global model.
  3. Identify the driving model and scenario. The IPCC notes that the quality of the driving model limits the result, and projections depend on the emission or concentration scenario. CORDEX itself leaves model selection to regional communities, so record which global models were chosen and why.
  4. Check the evaluation evidence. In CORDEX every group runs an ERA5-driven evaluation first. Gutowski et al. also note that observations used for evaluation can themselves contain errors.
  5. Match the variable and scale to the claim. Gutowski et al. state that the added value of downscaling depends on the variable, the statistics, and the temporal and spatial scales examined.
  6. Cite and archive the exact version. Use the persistent identifiers and archive records of the datasets and design documents you rely on; see citing data, data publication practices and the DOI definition.

Uncertainty is not one number. Gutowski et al. list its sources as systematic errors of the driving global models, errors of the regional models or empirical-statistical downscaling models, scenario and structural uncertainty, natural variability, internal variability of the regional model, and a sparse matrix of global-model and regional-model combinations. Natural variability becomes larger at finer scales, so a fine-grid result from a single run can overstate how certain a local number is.

Common misunderstandings

Downscaling corrects the global model’s errors

It does not. Gutowski et al. state that downscaling is not intended to correct biases in the large-scale circulations produced by the driving global models, summarising the point as “garbage in – garbage out”. Maraun et al. likewise write that no climate model or downscaling method can be expected to reproduce all aspects of climate perfectly, and that global-model error is an important source of regional uncertainty.

Higher resolution always means better information

The added value of downscaling is often debated, and Gutowski et al. say it does not always result in more valuable information. They also say the resource cost needs to be justified, particularly for very high-resolution (1-5 km) convection-permitting models. A finer grid should be treated as a hypothesis to test against the variable and scale you care about.

Statistical downscaling is a shortcut that avoids the driving model

Both methods start from larger-scale information. The IPCC states the driving model’s quality limits the downscaled information in all cases, whichever method is used.

The method is a neutral step

Maraun et al. note that the downscaling method itself is often a considerable source of uncertainty, which is why VALUE compares methods under common experiments rather than assuming any one is adequate.

Frequently asked questions

What is downscaling in climate science?

It is a method that derives local- to regional-scale information from larger-scale models or data analyses (IPCC AR6 glossary). The two main methods are dynamical and empirical/statistical downscaling.

What is the difference between dynamical and statistical downscaling?

Dynamical downscaling uses regional climate models, variable-resolution global models or high-resolution global models. Statistical downscaling uses observations to build statistical links between large-scale and local variables. They can also be combined.

Does downscaling fix bias in a global climate model?

No. Gutowski et al. (2016) state that downscaling is not intended to correct biases in the driving model’s large-scale circulation, and the IPCC says driving-model quality remains an important limitation.

What resolution does CORDEX-CMIP6 use?

The primary grid-spacing targets are 25 km and 12.5 km, with one common spacing per domain recommended. Finer spacing is allowed by agreement.

How do I validate a downscaling method?

The VALUE framework starts from the specific user problem, uses a validation tree to choose indices and measures, and runs experiments that isolate downscaling skill from error inherited from the global model.

Sources

Data: CORDEX-CMIP6 dynamical downscaling design parameters

Design parameters of the CORDEX experiment design for dynamical downscaling of CMIP6 (28 April 2025), compiled into one table from the design document and the Gutowski et al. 2016 CORDEX paper.

Period covered: 1950/2100.

CORDEX-CMIP6 dynamical downscaling design parameters
Design itemSpecificationSource
Standard domains14 continental-scale domains, including the Arctic and AntarcticCORDEX-CMIP6 experiment design
Primary grid-spacing targets25 km and 12.5 km; one common spacing per domain recommended; finer than 12.5 km by agreementCORDEX-CMIP6 experiment design
Minimum modelRegional atmosphere-land surface modelCORDEX-CMIP6 experiment design
Evaluation simulationDriven by ERA5 (1979-2024), one-year spin-up from 1979; ERA5.1 recommended for 2000-2006CORDEX-CMIP6 experiment design
Historical boundary conditionsCMIP6 historical, 1950-2014CORDEX-CMIP6 experiment design
Scenario forcingGreenhouse-gas forcing following SSP3-7.0 from 2015 onwardCORDEX-CMIP6 experiment design
Selection of global models to downscaleNo commonly accepted methodology; left to regional communitiesCORDEX-CMIP6 experiment design
Output distributionEarth System Grid Federation, CORDEX-CMIP6 archive specificationCORDEX-CMIP6 experiment design
Global forcing needed for dynamical downscalingSix-hourly global model fields, minimum 1950-2100Gutowski et al. 2016

Sources

Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.

Follow CASRAI

Research-administration guidance, standards updates and independent tool reviews.

Ask CASRAI · full access

Ask about Downscaling: Dynamical vs Statistical Methods Explained

Full access: large monthly usage, up to 10 source passages behind each answer, and MCP and API access for Claude Code, Codex and other CLIs. Cancel any time.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →