I detrended all the raw station data. That is, I estimate and subtract the trend for each station, leaving the mean over the observation period unchanged. So we have data which should return a zero trend. What do we get?
I looked at four cases:
- SO9 with 3 PC's (Eric Steig's original)
- SO9 with 5 PC's
- OLMC10 with RLS (regularized least squares)
- OLMC10 with E-W (eigenvalue weighting)
Well, none of the methods reported a zero trend, although the magnitudes were reduced relative to real data.
Update 2. As noted below, detrending the raw data is not a good idea - it is much better to calculate the detrend slope using anomalies. I did this, and found that it had little effect on the S09 results, but a big effect on the OLMC10 method. The RLS performance was now better. Qualitatively, though, the results are similar. The RLS method still gives a negative trend, by about as much as S09 is positive.
Here are the plots, using the same color range as in my last post:
S09 3 PCs
S09 5 PCs
Here are the new numerical trend averages in C/Decade:
This suggests also that S09 tends to return high trends - RLS low, to
The continent difference of
Update. As soon as I posted this, I realised that it would have been better to detrend anomalised data rather than the raw data. Estimating trends of seasonal series with missing values and irregular endpoints is noisy. I'll post revised results in about an hour when the computing is done.
Here is the table I calculated earlier detrending the raw data: