Analysing the ECMWF seasonal forecast fields: examples

The KNMI Climate Explorer allows you to correlate the 1991-2000 hindcast/forecast fields of 2m temperature, precipitation, sealevel pressure, z500 and 10m wind components with other climate data, such as the NCEP/NCAR reanalysis, CMAP (Xie&Arkin) precipitation estimates, ERS winds, and station data.

Some examples how one could proceed.

  1. Plot the data
  2. Pointwise correlations with other fields
  3. Fieldcorrelations with other fields
  4. Time series analysis of a regional average
  5. Search for downscaling relations

1. Plot the data

  1. Go to the home page and register,
  2. Proceed to the time series selection page,
  3. Click on "field" in the title to go to the field selection page,
  4. Jump or scroll down, select "ECMWF precipitation +3"
  5. Follow the link "plot this field",
  6. Fill out "1997", choose "Dec" and "3" and select "Take absolute anomalies", fill out the years "1991" to "1996", press "Plot"
This gives the DJF 97/98 +3month forecast anomalies relative to the ECMWF system-1 reference period, 1991-1996. Changing the scale to "-100" to "100" with a reversed colour scale "red-grey-blue" gives the figure below. Note that these are strict +3 months forecasts, i.e., the December values are from runs started in Aug/Sep, the January from runs started in Sep/Oct, and the February values from runs in Oct/Nov. There is no way to reproduce the plots on the ECMWF web site, which consist of one ensemble with the December forecast +3, the January one +4 and the February forecast +5 months.

Following the same procedure for observed precipitation gives a qualitative verfication.

2. Pointwise correlations with observations

  1. On the field selection page, select NCEP/NCAR "temperature" at "2m",
  2. Click on "correlate with another field (pointwise correlations)",
  3. Select the ECMWF +3 2m temperature, change "all" to "anomalies" to turn off the distinction between different seasons, and click on "correlate".
With a slightly clearer scale (from "-1" to "+1") one gets the figure below, showing mainly El Niño, its teleconnections and persistence.

3. Fieldcorrelations

  1. On the field selection page, select "CPC Merged Analysis of Precipitation",
  2. Follow the link "Correlate with another field (field correlations)",
  3. Select "ECMWF precipitation +3", press "correlate"
One sees that predictability is highest during an El Niño or La Niña. Of course this global field correlation is dominated by the tropics, where rainfall is highest. One can select part of the globe to zoom in on regions of interest.

The dependance on the state of the ENSO oscillation can be seen even more clearly in a scatterplot of this time series against the NINO3 index:

  1. Follow the link "Correlate with other time series",
  2. Select "NINO3", "together" and "Correlate".

The straight line is not the best fit, one can study the regions with NINO3<0 and NINO3>0 separately.

4. Average a region

  1. Select "CPC Merged Analysis of Precipitation",
  2. Follow the link "Average a region"
  3. Fill out the coordinates, e.g., "10" to "20"N, "-10" to "15"E and compute the time series.
  4. Do the same with the "ECMWF +3 precipitation"
  5. Click on "Correlate woth other time series"
  6. Select the CPC precipitation index you constructed in the first steps from the list of user-defined indices, and correlate. This gives a yearly cycle of correlations.
  7. The rain season looks promising, select "Jul", averegaing over "3" months to get JAS, and press "correlate" again.
You should now have the figure below (the years only show up in the postscript version). Note however the large (bootstrap) error margins due to the limited number of years.

You can also download the time series for further analysis back home. Unfortunately whole forecast fields cannot be downloaded.

5. Search for downscaling relations

Often the best predictor of a quantity of interest is not the corresponding forecast (if available), but another (set of) forecast parameters. This may be weather a grid point away (due to the coarse grid), or more large-scale features. One can try to identify these by correlating the 1991-2000 data of the predictand with the hindcast fields.

  1. Upload the time series of interest into the Climate Explorer (bottom of the time series selection page),
  2. Follow the link "correlate with a field",
  3. Try a few seasonal forecast fields and try to identify where the correlation is highest,
  4. Average these regions into time series,
  5. Construct a linear statistical forecast model
Of course the last step may be easier to perform with your own favorite software at home.

Good luck.

Geert Jan van Oldenborgh, KNMI, 10-sep-2001