N Cook:
|Monitoring 'progress' to 2100: Global Sea Level Rise from altimetry
|Latest output 07 November 2026 data to 26.729741 (28 Sep 2026).
|Indicial curve fit still the best fit and the century projection curve
|y = 2.490849 + 0.163822*x^1.237578
|y is global SLR in Aviso cm terms and x is year minus 2000, 6dp figures |retained for anyone else wishing to repeat/confirm these results.
|To 2100 processing all the data each time ,from 2003.0 to avoid |post-Pinatubo global geodata hiatus ,873 datapoints, SLR of 51.416 cm.
OK. I miss a lot since I went blind during the pandemic. But let's
see if I can check what's going on here.
I collect the Avisio data every month and whichever dataset I
have should be good enough to illustrate the problem with extrapolation
of noisy data. First of all I noticed the data I have is only ~2026 in the 12 cm range above their baseline, so it seems quite different to the sea heights you're using.
But anyway I took the 2000-2013 part and put it into a general curve fitter that minimises the rmse and it came up with an indicial curve (again)
much different from the one you cite but I forced it to find the
best fit but also made sure the power was >1 as in your model. I think that makes the extrapolation especially poor.
Now the "problem" comming with all extrapolation. You take some data and
try to figure out what things way out in the future will look like.
If you fit a curve then you must test that it actually is making a
good estimate of the future data by taking part of the data you have
and seeing how well the model it comes up with fits the part of the
data you did not show the curve fitter.
In this case we can use the model from the 2000-2013 data to fit
the 2014-2026 data. And while the model fits the 2000-2013 great,
it "makes a little mistake" and latches onto something that makes
its predictions increasingly wrong the further out your get.
This is a warning that if you make a "best fit" model using all the data
you have 2000-2026 then expecting it to work at 2050 or 2100 is
going to be disappointing.
Here's what the procedure I used did as a plot: <http://kym.massbus.org/graphs/indicialmodel.gif>.
You can see if fits the first 1/2 of the data fine. But it unfortunately latches onto the "wrong trend" and more or less continues on that track.
The data just ~2013 seems to jump to new tram tracks making the real
data post 2013 now accelerate above the trend the model predicts.
So the data-sciency way to make sure you can predict something
"well enough" way out into the future is to find some curve that
you can prove extrapolates from e.g. 1/2 the data you have to
the other 1/2 of the data you have. If you can do that you can
"hope" (that's about all you can do :) the same procedure applied to
all the data you have will "kinda work" for data around that same
timeframe out into the future and maybe even further.
| Sysop: | Amessyroom |
|---|---|
| Location: | Fayetteville, NC |
| Users: | 76 |
| Nodes: | 6 (0 / 6) |
| Uptime: | 61:00:46 |
| Calls: | 1,265 |
| Calls today: | 1 |
| Files: | 1,362 |
| D/L today: |
4 files (9,632K bytes) |
| Messages: | 286,219 |