Seasonal forecast data from the UK Met Office

A user asked this question and it may be useful information for other users:

"For the UK Met Office seasonal forecast single level dataset on the CDS, there are three systems (12, 13 and 14) which one should I use?

For example the system 12 has forecast years between 2017-2018 whilst system 14 only for the year 2019.

If I want the whole time series of hindcasts and 2017-2019 forecasts should I combine the systems or they are different?"



In principle, as a general rule, I would say that it is NOT a good idea to combine data coming from different systems. It might depend on what is it meant by "combine" and "complimentary data", but if I understood it properly, it was suggested to pool them all together and that introduces interesting questions that need to be tackled:

  • how do you treat hindcast dates where more than one system is available?
  • which hindcasts (system) do you use to bias-correct each real-time forecast (different systems)?

As it is explained in footnote (d) of the table at the bottom of this page in the documentation (, due to the characteristics of the UK MetOffice seasonal forecast system design, the use of the keyword "system" has a different meaning than in other contributors datasets.

The UK MetOffice produces their hindcasts "on-the-fly", meaning that every time they issue a new real-time forecast, they also produce the full set of hindcast data relevant for that forecast. That approach allows them to introduce more frequent changes in their forecasting system (in comparison to those systems using a fixed set of hindcasts produced as a one-off). In that context, the "system" keyword used for UK MetOffice in our GRIB files archived in MARS is changed once per year, so the new hindcasts are not messed up with those produced the previous year for the same start dates. The implications of this are the following:

  1. during the lifetime of a given value of keyword "system", changes to the forecasting system might have happened.
  2. a change in the value of "system" keyword is not necessarily aligned with an underlying change in the forecasting system.

All those design details make impossible to have a continuous record of data for a given UK MetOffice "system" (continuous, in the sense of no gaps between the end of the hindcast period and the period for which that "system" produced real-time forecasts).

Due to all of that, my recommendation would be to use the most recent set of hindcasts and, where needed, their associated real-time forecasts.
At the time of writing this comment (September 2019), this can be translated into two possible approaches:
(a) Use system=14 (hindcasts available from May to Oct, soon Nov) and system=13 (from Nov or soon Dec to April); the real-time forecast period covered with that would be Nov 2018 to Sep 2019
(b) Use system=13 (you have a full set of 12 start dates of that one); the real-time forecast period would be May 2018 to Apr 2019