Creation of ER-M-Climate

The ER-M-Climate is derived from a set of extended range re-forecasts created using the same calendar start dates over several years for data times either side of the time of the extended ensemble run itself.  The re-forecast runs are at the same resolution as the extended medium range run itself and run over the 46-day extended range ENS period. 

There is merit in examining the real-time performance of a forecasting system.   But the sample sizes created for one system are far too small to conclude anything about its true performance levels.  Re-forecasts are used to increase the available data to produce a model climate.   The results of forecast system may be compared with this model climate.

Re-forecasts are a fundamental component of all seasonal forecasting system; they have two applications:

Selection of extended range re-forecasts

The set of re-forecasts is based on using the three consecutive dates surrounding the day and month of the extended ENS run in question.  Re-forecasts are created using the same calendar start dates for each of the last 20 years.  

The set of re-forecasts is made up from:

In total, each set of re-forecasts consists of 20 years x 3 runs x 11 ensemble members = 660 re-forecast values.  These are available for each forecast parameter, forecast lead-time, calendar start date, location, at forecast intervals of 6 hours.  These are used to define the ER-M-climate.

A lower number of re-forecasts than for evaluating M-climate is justified because the tails are less important and should not be so prone to having a reduced sample size.

The ER-M-climate is used in association with the extended range ensemble forecast:

Different reference periods for M-Climate and ER-M-Climate

ECMWF uses different reference periods but essentially the same re-forecast runs to build the M-Climate and the ER-M-Climate.   The key difference is that those runs are grouped and used in different ways:  


Note before Cy41r1 in spring 2015, the M-climate was constructed from only 500 re-forecasts was more prone to sampling errors and as a result.