CDO

Hi,

Are the output files containing correct CRPS calculation or the number of timesteps in rfile and ifiles should be the same?

Yes, it must be same.

The overlapping time-steps can be adjusted by setting the environment variable "export SKIP_SAME_TIME=1", see the thread/post by Uwe here https://code.zmaw.de/boards/1/topics/1134#message-1136

Cheers,
J.

Thank you for your answer. But skipping the overlapping time-steps would remove information from the ifiles because although the time is the same, the data contained is different. For example one time-step for one date and time may include h+3 data and the next time-step (but same time) includes h+27 data from the previous day.
regards,
Juanma

Hi,

okay, I thought it just a repetition due to merging. Nevertheless, the time steps of the reference, and ensemble(s) MUST be same.

because although the time is the same, the data contained is different. For example one time-step for one date and time may include h+3 data and the next time-step (but same time) includes h+27 data from the previous day.

but the forecast reference time "h", is different and CDO won't automatically recognize the forecast reference time and compare it with appropriate verification time. One way is to do the CRPS for each set of forecast ensemble(s) starting at a particular initial time, and finally take the average CRPS across difference sets

OR

you just repeat your reference data at that particular verification time, for example in your member file you have two data sets for "2010-11-30 00:00:00", and only one in reffile (obviously), so

cdo mergetime -seldate,2010-11-30T00:00:00 reffile reffile new_merge

like that for all the "repeating" forecast hours in the member ensemble - that you have to do it manually one by one

Cheers,
J.

Thank you,
I think the second option that you suggest is the best and easier to implement. Please confirm if I have understood it: You mean that I should change the reffil in order to repeat the analysis data for the overlapping period times: 2010-11-30 00:00:00 2010-11-30 00:00:00 2010-11-30 03:00:00 2010-11-30 03:00:00 2010-11-30 06:00:00 2010-11-30 06:00:00 2010-11-30 09:00:00 2010-11-30 09:00:00 2010-11-30 12:00:00 2010-11-30 12:00:00 and leave the rest as they are (just one for each analysis time). Afterward the reffile with the analysis and ensemble files will have exactly the same number of fields. So I understand that the algorithm developed to calculate CRPS works picking up sucesive data pairs from reffiles and the ensemble files. Am I correct?
Thanks a lot for your help
Juanma

Hi,

You mean that I should change the reffil in order to repeat the analysis data for the overlapping period times ...
and leave the rest as they are (just one for each analysis time). Afterward the reffile with the analysis
and ensemble files will have exactly the same number of fields.

Yes, exactly!

Don't forget to double-check the final reference data! (cdo info -seldate,... or cdo -info -seltimestep.. etc or simpe stat operation sutraction/addition etc)

So I understand that the algorithm developed to calculate CRPS works picking up sucesive data pairs from reffiles and the ensemble files.

Yes, CRPS values at each time step - a CRPS time series. In most stat operations CDO (like most stat analysis tools) won't double check what the user gave as input, it only check the spatial/temporal dimensions (In your case, it was the time dimension).

Cheers,
J.

Thank you very much for your quick answers!!