GNOME Bugzilla – Bug 385467
implement double exponentional smoothing
Last modified: 2008-09-26 08:50:23 UTC
In gnumeric cvs, there is a nice new feature called "exponential smoothing". It allows to smooth a noisy curve. However, what is currently implemented is a simple exponentional smoothing, which has the side effect of modifying the phase of a curve when there is a trend. This make the exponentional smoothing useless in a lot of cases. To prevent this, there is the so-called double exponentional smoothing. It is explained on this page: http://www.toyproblems.org/probs/p02/ Thank you.
thanks for the link.
After reading the docs, I conclude that what I implemented as exponential smoothing is not really simple exponential smoothing but an extension of the concept (it uses the abscissa and the data might be unordered). The question is what should I implement now? - something similar to what we already have, but more like double exponential smoothing? - or the classic simple and double exponential smoothing formula (and also triple exponential smoothing by the way)?
Hi Jean, I have not further investigated this problem (and I have to leave for a few weeks with http access), but I would say that: - it is better to have something that works on non-ordered series - the smoothing should be symetrical: reversing the time (the abscissas) should not change the result Cheers, Frederic
exponential smoothing is not symmetrical. You take into account only what happened before, so if you revert the time, the data used are obviously not the same. Of course if there is a trend, it must be similar in both cases.
Andreas, I think it should be implemented more or less the way I did it for simple exponential smoothing (which allows non ordered data). I already have some code, and was planning to finish that soon, but of course, if you prefer, you can do it.
The exponential smoothing tool is one of the few left that calculates values only only rather than inserting formulas like most of the other analysis tools. So I am basically set to rewrite that tool completely.
I have added Holt's trend corrected exponential smoothing (which is what is usually meant with double exponentail smoothing). This problem has been fixed in the development version. The fix will be available in the next major software release. Thank you for your bug report.
Excellent, thank you very much for this! Frederic