ENH: resample().interpolate()

[jorisvandenbossche]

From #12449 (comment)

When downsampling on a Resampler object, you now have different fillna methods to fill the NaNs (or asfreq for a plain reindex like operation without NaN filling). I can possibly make sense to also have interpolate available to fill the missing values directly (instead of first calling mean/asfreq)

In [20]: df = pd.DataFrame(data=[1,3], index=[dt.timedelta(), dt.timedelta(minut
es=3)])

In [21]: df
Out[21]:
          0
00:00:00  1
00:03:00  3

In [22]: df.resample('1T').mean().interpolate('linear')
Out[22]:
                 0
00:00:00  1.000000
00:01:00  1.666667
00:02:00  2.333333
00:03:00  3.000000

Should df.resample('1T').interpolate('linear') give the same result?

[benoit9126]

Hi!
Finally, I found some time to create what could be considered as a pull request for #12925. Unfortunately, in my additional tests, I face the problem raised by @jorisvandenbossche with .asfreq in #12449.

In[39]: df.resample('1T').ffill()
Out[39]: 
          0
00:00:00  1
00:01:00  1
00:02:00  1
In[40]: df.resample('1T').bfill()
Out[40]: 
          0
00:00:00  1
00:01:00  3
00:02:00  3

It seams that .mean() is the only method which retrieves 4 values in this example and of course, the additional interpolate method of the Resampler has the same problem...

Should I propose my PR with a test which fails because of this bug? or should I remove one of my test for the moment?

[jreback]

#12928 fixed that. make sure you are up to date.

[benoit9126]

Are you sure it really works??

In[2]: import pandas as pd
In[3]: pd.__version__
Out[3]: '0.18.0+142.g92b5322'

So the latest version. (10 commits above #12928)

Here an example with timedeltas:

In[4]: df = pd.DataFrame(data=[[1, 3], [-5, 10], [0, 0]],index=pd.timedelta_range('00:00:00', '00:10:00', freq='5T'))
In[5]: df
Out[5]: 
          0   1
00:00:00  1   3
00:05:00 -5  10
00:10:00  0   0
In[6]: df.resample('1T').mean()
Out[6]: 
            0     1
00:00:00  1.0   3.0
00:01:00  NaN   NaN
00:02:00  NaN   NaN
00:03:00  NaN   NaN
00:04:00  NaN   NaN
00:05:00 -5.0  10.0
00:06:00  NaN   NaN
00:07:00  NaN   NaN
00:08:00  NaN   NaN
00:09:00  NaN   NaN
00:10:00  0.0   0.0
In[7]: df.resample('1T').asfreq()
Out[7]: 
            0     1
00:00:00  1.0   3.0
00:01:00  NaN   NaN
00:02:00  NaN   NaN
00:03:00  NaN   NaN
00:04:00  NaN   NaN
00:05:00 -5.0  10.0
00:06:00  NaN   NaN
00:07:00  NaN   NaN
00:08:00  NaN   NaN
00:09:00  NaN   NaN

And here the example with datetimes:

In[8]: df = pd.DataFrame(data=[[1, 3],[-5,10]], index=pd.date_range('22/4/2016 00:00:00', '22/4/2016 00:05:00', freq='5T'))
In[9]: df
Out[9]: 
                     0   1
2016-04-22 00:00:00  1   3
2016-04-22 00:05:00 -5  10

In[11]: df.resample('1T').mean()
Out[11]: 
                       0     1
2016-04-22 00:00:00  1.0   3.0
2016-04-22 00:01:00  NaN   NaN
2016-04-22 00:02:00  NaN   NaN
2016-04-22 00:03:00  NaN   NaN
2016-04-22 00:04:00  NaN   NaN
2016-04-22 00:05:00 -5.0  10.0
In[12]: df.resample('1T').asfreq()
Out[12]: 
                       0     1
2016-04-22 00:00:00  1.0   3.0
2016-04-22 00:01:00  NaN   NaN
2016-04-22 00:02:00  NaN   NaN
2016-04-22 00:03:00  NaN   NaN
2016-04-22 00:04:00  NaN   NaN
2016-04-22 00:05:00 -5.0  10.0

The behaviour is still not the same between .mean() and .asfreq() (and some other operations .bfill(), etc.) for timedeltas.

[jreback]

ok, so soln is not complete. seeing if I can fix.

[benoit9126]

Thanks a lot.