Re: [Gnuplot-info] Fit trouble when I use xdata time (fit isn't so good)

[Hans-Bernhard B. <HBB...@t-...>]

Guilherme Poleszuk wrote:

> I believe that are a problem with "fit" function when I use xdata time, 
> because X values are converted to second since 2000/01/01 ( I think! ) and 
> this number get big compared to Y values... 

Exactly.  Fit is somewhat allergic to parameter values of wildly 
different magnitude.  The issue is not unique to time-date axes, but 
those do have a tendency to get hit by it more often than others because 
they tend to pick up big figures, like the number of seconds in a day 
(or a year, even), as factors that end up in the parameters.

The solution is to shift or scale the problem to get the expected 
parameters into similar magnitude ranges.  E.g. it might help to define 
your model function to take days instead of seconds as its argument

	fit f(x/86400) 'data' u 1:4 via a, b

or (like Thomas suggested already), shift the zero point of your x axis.

Note that the usual simple forms of the straight line equation

	y = m*x + y0
	y = m*(x -x0)

may suffer from ginormous extrapolation from a very short piece of curve 
all the way to the point of x=0 (or y=0), causing parameters like x0 or 
y0 to be entirely dominated by rounding error.