[SciPy-Dev] scipy.stats.distributions: note on initial parameters for fitting the beta distribution

josef.pktd@gmai... josef.pktd@gmai...
Mon Nov 1 22:13:51 CDT 2010


On Sun, Oct 31, 2010 at 9:34 AM, James Phillips <zunzun@zunzun.com> wrote:
> File attached.
>
> On Sun, Oct 31, 2010 at 8:33 AM, James Phillips <zunzun@zunzun.com> wrote:
>> Here is a more polished and quite smaller version of your example file
>> matchdist.py that uses either nnlf or residuals for ranking, and
>> includes checks for NaN, +inf and -inf.  I think this has all of the
>> logic and range checks that it needs.

I tried out both your scripts during the weekend but didn't get around
to replying. Most distributions work pretty fast but there are still a
few unsuccesful time wasters in there. For example, ksone is also
mainly a distribution for a statistical test, and in my run took a
long time without a successful fit.

I'm not sure how nnlf woill work as selection criterium for the
distributions, and similarly the residual sum of squares might not be
a good or robust criterium, for example with heavy tailed
distributions. But that's just a guess, the only (commercial) package
that I looked at, offered the choice between Kolmogorov-Smirnov,
Anderson-Darling and 2 chisquare tests (equal-spaced and equal
probability) as distance or goodness-of-fit measure. (Entropy would be
another criterium, but I haven't seen it yet for selecting the
distribution.)

Your version also will help to narrow down what might be good starting
values, eventually I would prefer to hardcode (optional) distribution
specific _start_values.

Your second script (after dropping the failures) has 10 distributions
fewer than the first script (70 instead of 80), so there is still some
distribution specific work left.

Josef

>>
>>     James
>>
>> 2010/10/30 James Phillips <zunzun@zunzun.com>:
>>> I'll parallelize this code and make a few more tweaks, and then add it
>>> to my web site.
>>
>
> _______________________________________________
> SciPy-Dev mailing list
> SciPy-Dev@scipy.org
> http://mail.scipy.org/mailman/listinfo/scipy-dev
>
>


More information about the SciPy-Dev mailing list