Met Office Blended Probabilistic Forecast – UK gridded percentiles

This product provides gridded percentile weather forecasts. The grid resolution is approximately 2km and covers the UK and parts of Western Europe. It is produced by the Met Office IMPROVER Blended Probabilistic Forecast system. It is available in NetCDF format.

Blended Probabilistic Forecast data is derived from the Met Office's operational NWP (Numerical Weather Prediction) ensembles and nowcasts. To give more reliable predictions, these are then blended and calibrated using the IMPROVER pipeline, and verified using spread–skill and reliability checks.

This is 1 of 8 Blended Probabilistic Forecast products published by the Met Office on the Registry of Open Data on AWS. Data is available for the Global and UK domains, as gridded and spot (site-specific), and represented as percentiles and probabilities.

This info is correct as of April 2026, but some things (like the number of sites, parameters and timesteps) may change in future.

How percentiles work

Ensemble forecasts show a range of possible weather outcomes. However, some users may find it more useful to see ensemble forecasts presented as percentiles, particularly when they want to see where each member of an ensemble sits within the full range of possible outcomes.

Percentiles are generated from an ensemble forecast by first sorting all the members of that ensemble, for example from the coldest temperature to the warmest. To then identify a particular percentile forecast (e.g. the 10th percentile), we find the temperature at which 10% of the ensemble members predict colder conditions. As only 10% of the ensemble members are predicting a lower temperature than the 10th percentile forecast, this indicates that it is giving a relatively low forecast of temperature. Similarly, 90% of ensemble members would predict a lower temperature than the 90th percentile forecast, indicating that it is giving a relatively high forecast of temperature.

Precipitation percentiles should be used cautiously

Precipitation percentiles are potentially useful for experts. But we don't recommend that most people use them, as they are hard to interpret.

Precipitation is spatially noisy (i.e. it can vary a lot over small distances), especially when it is showery. This means that the individual ensemble forecasts that the percentiles are generated from are likely to have their heaviest precipitation positioned in different places. As a result, a high percentile (e.g. 95th) will pick up the high values from all those different locations and make it appear that heavy rain could occur over a wider area than is physically plausible. In other words, the spatial extent of the precipitation when using a high percentile is not physically realistic. High percentiles can be very useful for finding out what the values at the higher end may be, but not how they are spatially organised.

Likewise, the spatial extent of the precipitation will be greatly reduced for the lower percentiles. If there are differences between the ensemble forecasts about where it will rain or not, it is possible that a low percentile precipitation field may show zero precipitation everywhere. Again, that is potentially misleading because it suggests it might be dry everywhere, which is not what the individual forecasts are necessarily saying. It is better to view different percentile maps together (5th, 50th, 95th) to get a more informative impression.

Even the 50th percentile can be misleading as, by definition, it will never include the highest predicted values. Nor is there any guarantee that the peaks in the 50th percentile grid will align with the peaks in the 95th percentile grid.

About the grid

The grid resolution is approximately 2km and covers the UK and parts of Western Europe.

Numerical weather prediction (NWP) models generate forecasts for each grid point within a geographical area of interest. Each of these gridded forecasts corresponds to a particular diagnostic (e.g. precipitation rate) at a particular time. IMPROVER then takes an ensemble of these gridded forecasts and applies post-processing techniques to enhance them and represent them probabilistically. The resulting grid of values represents probabilities of exceeding or falling below a particular threshold.

Parameters and timesteps

There are 44 weather parameters available including:

• Cloud
• Temperature
• Pressure
• Humidity
• Visibility
• Precipitation rate and accumulations (see note above about care required with use)
• UV
• Wind

For most parameters, the following timesteps are available:

• Every hour from 0 to 120 hours
• Every 3 hours from 123 to 186 hours

However, timesteps vary significantly for some parameters. Check the parameter documentation  for more details.

Latency

Data is made available shortly after the model blend time.

Archive length

Data is available for the past 30 days.

Business needs

This product supports risk-based decision-making by providing uncertainty ranges rather than single deterministic values. Typical uses include:

• assessing uncertainty for operational planning
• evaluating weather-related risk thresholds
• deriving deterministic products (e.g. 50th percentile) from probabilistic outputs

Gridded forecasts show how a diagnostic varies spatially across a domain at a given time. By using a time series of gridded fields, you can determine how a weather diagnostic is expected to evolve across a geographic area.

If you need a forecast for a specific location, a spot forecast may suit your needs better than gridded data. Spot Blended Probabilistic Forecasts are also available as percentiles and probabilities for both the UK and Global domains.