Pandas: Interpolate NaT

[scls19fr]

Hello,

interpolate doesn't work with NaT
see http://stackoverflow.com/questions/33921795/fill-timestamp-nat-with-a-linear-interpolation/33922824#33922824

Here is a trivial example to show the situation:

s = pd.Series(pd.date_range('2015-01-01' , '2015-01-30'), name='t')

s[3], s[4], s[5] = pd.NaT, pd.NaT, pd.NaT
s[13], s[14], s[15] = pd.NaT, pd.NaT, pd.NaT
print(s)

0    2015-01-01
1    2015-01-02
2    2015-01-03
3           NaT
4           NaT
5           NaT
6    2015-01-07
7    2015-01-08
8    2015-01-09
9    2015-01-10
10   2015-01-11
11   2015-01-12
12   2015-01-13
13          NaT
14          NaT
15          NaT
16   2015-01-17
17   2015-01-18
18   2015-01-19
19   2015-01-20
20   2015-01-21
21   2015-01-22
22   2015-01-23
23   2015-01-24
24   2015-01-25
25   2015-01-26
26   2015-01-27
27   2015-01-28
28   2015-01-29
29   2015-01-30
Name: t, dtype: datetime64[ns]

print(s.interpolate())
0    2015-01-01
1    2015-01-02
2    2015-01-03
3           NaT
4           NaT
5           NaT
6    2015-01-07
7    2015-01-08
8    2015-01-09
9    2015-01-10
10   2015-01-11
11   2015-01-12
12   2015-01-13
13          NaT
14          NaT
15          NaT
16   2015-01-17
17   2015-01-18
18   2015-01-19
19   2015-01-20
20   2015-01-21
21   2015-01-22
22   2015-01-23
23   2015-01-24
24   2015-01-25
25   2015-01-26
26   2015-01-27
27   2015-01-28
28   2015-01-29
29   2015-01-30
Name: t, dtype: datetime64[ns]

assert s.interpolate().isnull().sum() == 0
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
<ipython-input-150-8a59e397a174> in <module>()
----> 1 assert s.interpolate().isnull().sum() == 0

AssertionError:

Kind regards

[jreback]

this is not implemented ATM on datetimes. pull-requests are welcome.

[scls19fr]

What is your opinion about the last solution proposed by CT Zhu on StackOverflow ?

df.ix[df.t.isnull(), 't'] = pd.to_datetime(pd.to_numeric(df.t).interpolate())[df.t.isnull()]

isn't there a method to support NaN with integers without converting to float (which lead to precision issue) ?

[scls19fr]

Shouldn't we look for example to

np.int64(pd.NaT)

which is -9223372036854775808

[axelv]

I have the impression that interpolating NaT is still not possible in v20.3.
Any updates on this issue?

[jreback]

This is not very hard to actually do directly (and what .interpolate() should basically do, PRs welcome)

In [12]: s2 = s.astype('i8').astype('f8')

In [13]: s2[s.isnull()] = np.nan

In [14]: pd.to_datetime(s2.interpolate())
Out[14]: 
0    2015-01-01
1    2015-01-02
2    2015-01-03
3    2015-01-04
4    2015-01-05
5    2015-01-06
6    2015-01-07
7    2015-01-08
8    2015-01-09
9    2015-01-10
10   2015-01-11
11   2015-01-12
12   2015-01-13
13   2015-01-14
14   2015-01-15
15   2015-01-16
16   2015-01-17
17   2015-01-18
18   2015-01-19
19   2015-01-20
20   2015-01-21
21   2015-01-22
22   2015-01-23
23   2015-01-24
24   2015-01-25
25   2015-01-26
26   2015-01-27
27   2015-01-28
28   2015-01-29
29   2015-01-30
Name: t, dtype: datetime64[ns]

[scls19fr]

@rinoc did some work on this issue in https://github.com/rinoc/pandas/commit/e77e4c8566db68c0ec144f9aeb01dc5225c971d6
But no PR have been send.
Any news?

[mroeschke]

Looks to be a duplicate of https://github.com/pandas-dev/pandas/issues/11312