How to set series frequency?

[Peque]

Having two time series like:

import pandas
from numpy import nan


# 1-minute period
index = pandas.date_range('1/1/2010', periods=8, freq='T')
series1 = pandas.Series([1., nan, nan, nan, nan, 2., nan, nan], index=index)

# 5-minutes period
index = pandas.date_range('1/1/2010', periods=2, freq='5T')
series5 = pandas.Series([1., 2.], index=index)

One of them is 1-minute based, the other is 5-minutes based. They both have the same values (excluding nan) at the same timestamps.

I do not want to keep nan values, so I drop them. However, dropping them also drops the frequency information from the DatetimeIndex:

series1 = series1.dropna()
print(series1.index.freq)

Now, I would really like to be able to differenciate two series like these: same timestamps, same values, but originating from different base frequencies.

In order to do so, I was able to set the frequency back "by hand":

from pandas.tseries.frequencies import to_offset


series1.index.freq = to_offset('T')

But since 0.23 it seems I can no longer do that:

ValueError: Inferred frequency None from passed values does not conform to passed frequency T

What can I do now to keep that frequency information?

Also, is it incorrect to set a frequency "by hand" for an index that has that frequency but some missing values?

[gfyoung]

cc @jreback @mroeschke

[mroeschke]

Well freq is not a label but rather describes the spacing between the timeseries. So once series1 drops the nans the timeseries no longer conforms to 'T' frequency:

In [4]: series1
Out[4]:
2010-01-01 00:00:00    1.0
2010-01-01 00:05:00    2.0
dtype: float64

So I'd argue the behavior before was a bug. If you want to preserve original frequency as a label, just use the name attribute on the series.

In [9]: index = pandas.date_range('1/1/2010', periods=8, freq='T')
   ...: series1 = pandas.Series([1., nan, nan, nan, nan, 2., nan, nan], index=index, name='T')
In [11]: series1.dropna()
Out[11]:
2010-01-01 00:00:00    1.0
2010-01-01 00:05:00    2.0
Name: T, dtype: float64

[jschendel]

So I'd argue the behavior before was a bug.

xref #20678 (specifically item 2)