Support mean() for Timestamps.

[sorenwacker]

Fitting a linear model to Timestamps currently does not work out of the box.
and pandas does not support the ols function any further. When using e.g. seaborn to make a fit it throws an error.

TypeError: reduction operation 'mean' not allowed for this dtype

I wonder whether the dtype can be change so that mean is supported? And whether this would make it compatible with other packages, such as scikit-learn, seaborn, etc. Also, the mean of some dates could be useful for other use cases. It does make sense to calculate the mean of some dates doesn't it?

[sorenwacker]

Note, how to fit a model using datetimes/timestamps comes up frequently on stackoverflow in different variations.

[gfyoung]

cc @jreback @jorisvandenbossche

[jreback]

i believe we have an issue for this - can you see if this is duplicate

[gfyoung]

@jreback : It's similar to #17382, but that's for Timedelta instead of Timestamp.

[mroeschke]

Some thoughts:

1. Would this change describe()? Currently we provide top, freq, first, last (categorical summary) for Timestamp data.

2. What about other ops? (std, median, quantile, etc.) For example std of Timestamp data should probably return a Timedelta but how about skew and kurtosis (or do they not make sense)?

[sorenwacker]

I agree, median, std, quantiles etc would be very useful too. Also, subtraction of dates would be nice. Why are they called first and last if it could be min and max as well?

[jbrockmendel]

Is there anything left to do here?