[Numpy-discussion] number datetime64 dtypes

[Eric Wieser]

It's interesting to confirm that people are aware of this syntax!

This is intended but perhaps not useful behavior.

`datetime64[15D]` is a type that stores dates by the nearest date that is a
multiple of 15 days from the unix epoch.
Arguably there isn't a situation where using `15D` makes a whole lot of
sense, but the generalization is useful - `datetime64[15m]` stores dates
rounded to the nearest quarter hour, which is somewhat sensible.

Perhaps we should have added support for a custom epoch, which would make
your problem go away...

On Thu, 10 Sep 2020 at 18:43, Dr. Mark Alexander Mikofski PhD <
mikofski at berkeley.edu> wrote:

> Hi,
>
> Thank you for your time.
>
> A colleague asked me about creating a range of numpy datetime64 at 15-day
> increments.
>
> This works:
>
> np.arange(np.datetime64('2008-04-01'), np.datetime64('2020-09-01'), np.timedelta64(15, 'D'))
>
> but then they also showed me this, which leads to some very strange
> responses:
>
> np.arange(np.datetime64('2008-04-01'), np.datetime64('2020-09-01'),
> dtype="datetime64[15D]")
> Out[50]:
> array(['2008-03-27', '2008-04-11', '2008-04-26', '2008-05-11',
>        '2008-05-26', '2008-06-10', '2008-06-25', '2008-07-10',
> ...
>        '2020-05-23', '2020-06-07', '2020-06-22', '2020-07-07',
>        '2020-07-22', '2020-08-06'], dtype='datetime64[15D]')
>
> See how the 1st day is March 27th?
>
> I couldn't find a reference to this dtype ( "datetime64[15D]" ) in the
> numpy docs, but I think it's a common pattern in Pandas, that is using a
> number to get an increment of the frequency, for example "5T" is 5-minutes,
> etc.
>
> There is a reference to using arange with dtype on the datetimes &
> timedelta doc page () but the datetime is 1-day or  "datetime64[D]"
>
> Is this the intended outcome? Or is it a side effect?
>
> I wonder if others have tried to adapt Pandas patterns to Numpy datetimes,
> and if it's an issue for anyone else.
>
> I've advised my colleague not to use Numpy datetimes like this, assuming
> based on the docs that Pandas-style offsets do not translate into Numpy
> style datetimes.
>
> thanks!
>
> --
> Mark Mikofski, PhD (2005)
> *Fiat Lux*
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at python.org
> https://mail.python.org/mailman/listinfo/numpy-discussion
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20200910/da3bb220/attachment.html>