So with the ESRL data, I subtract the mean of the respective months of 1994-2013 from the overall monthly totals, and compare with the monthly averages I post. Here is a table of discrepancies for 2015, in the monthly anomalies, in °C
Mostly it is accurate to 3 sig figs. The exception is February, which also stands out in 2014. I think the reason is the treatment of leap years. With ESRL I subtracted the means of all Februaries in the base from the monthly means, while in my posting method, I form anomalies for each date first, abd then average to get the month. So in effect, for 2015 I compare with the average of all periods Feb1-28. Both ways are reasonable, and the difference is small.
AFAIK, there is no published global average of daily anomalies, so I will keep doing it the way I do. But it is useful to have the check of the monthly averages, and also to know that there is such little difference between surface and sig995.
I'm currently dealing with change of year issues - I'm restructuring the whole posting calculation to make it less year-dependent. Numbers for 2016 should appear soon.
There is really little difference.ReplyDelete
As for difference between base period, if you take 1994-2013, as you do, you have in december the greatest anomaly for NCEP/NCAR. But if you take 1951-1980, as GISS, you get october warmer than december, by far (+1,12°C vs 0,9°C). I wonder how you calculate you GISS adjusted. It seems you add the average difference between the two datasets. But why didn't you take NCEP/NCAR 1951-1980 as a base for GISS adjusted, I don't know. Anyway, you would not have more accurate results if you did.
"But if you take 1951-1980, as GISS, you get october warmer than december"
Yes, I'm well aware of the month variation issues. I wrote a post about it here, which included a table of offsets to convert from one period to another. For the GISS numbers, I used those - so the conversion only uses the same months measured with GISS and NCEP. That is, for May I add the offset of difference GISS May average from 1994-2013 and 1951-80.
I used 1994-2013, because that was the period of good data for NCEP/NCAR (explained here). Before that, there are missing patches, and it seemed generally ragged, as you might expect with retrospective data assimilation.
Thanks. Indeed, base 51-80 shows huge deviation around the mean for ncep. À bigger drop from oct to Nov 2015 that with base 94-2013.ReplyDelete