On 26/02/2019 10:15, Martin Brown wrote:
On 26/02/2019 09:57, N_Cook wrote:
y=2.483612 + 0.053029*x^1.575023

Just because you can obtain a model fit with 6 significant figures shown
in each coefficient doesn't mean that they are right.

I'd be very surprised if anything more than the first two leading digits
were even in the right ballpark.

Instead of using R^2 as a measure of goodness of fit why don't you try
and do it using the actual residuals of the data and an estimate of the
noise to compute chi-squared. Then you can do a proper analysis and see
just how uncertain each of your coefficients are. IOW how well defined
each of them is with the others sat on their respective peak value.


Its too early to have any confidence in this sort of stuff. I'm
interested how the estimations vary over time, uing the same baseload of
data , 2003 to early 2018 , and seeing how updates vary the output using
the same treatment each time.
If only the likes of Aviso would move away from linear "fit" I would not
feel being obliged to do any of this. I still have not found anyone
publishing similar results, using the Jason dataset (I have little
confidence in the Attica/Saral concept of calibration).
So far its been consistent ranking of curve types,
linear,exponential,quadratic, indicial, will that remain so?
After all they have big machines they can plug in full datasets and
churn away.
Incidentally has anyone found where the Jason-3 data is supposed to be
on the copernicus.eu site?