ENH: resample(..., base='start') for automaticly determining base.

[dimbeto]

(edit):

import pandas as pd
import datetime as dt
import numpy as np

datetime_start = dt.datetime(2014, 9, 1, 9, 30)
datetime_end = dt.datetime(2014, 9, 1, 16, 0)

tt = pd.date_range(datetime_start, datetime_end, freq='1Min')
df = pd.DataFrame(np.arange(len(tt)), index=tt, columns=['A'])

The first item is at 9:30, which is not divisible by 8 minutes. We'd like df.resample('8T', base='start').first() to be equivalent to df.resample('8T', base=2).

In [13]: df.resample("8T", base=2).first()
Out[13]:
                       A
2014-09-01 09:30:00    0
2014-09-01 09:38:00    8
2014-09-01 09:46:00   16
2014-09-01 09:54:00   24
2014-09-01 10:02:00   32

---

It seems that for 1Min bar data, resample() with sampling frequency of any multiple of 8 has a bug. The code below illustrates the bug when resampling is done at [3, 5, 6, 8, 16] Min. For both 3 and 5 frequency, the first entry of the resampled dataframe index starts at the base timestamp (9:30 in this case) while for frequencies 8 and 16, the resampled index starts at 9:26 and 9:18 respectively.

import pandas as pd
import datetime as dt
import numpy as np

datetime_start = dt.datetime(2014, 9, 1, 9, 30)
datetime_end = dt.datetime(2014, 9, 1, 16, 0)

tt = pd.date_range(datetime_start, datetime_end, freq='1Min')
df = pd.DataFrame(np.arange(len(tt)), index=tt, columns=['A'])

for freq in [3, 5, 6, 8, 16]:
    print(freq)
    print(df.resample(str(freq) + 'Min', how='first', base=30).head(2))

produces the following output:

3
                     A
2014-09-01 09:30:00  0
2014-09-01 09:33:00  3
5
                     A
2014-09-01 09:30:00  0
2014-09-01 09:35:00  5
6
                     A
2014-09-01 09:30:00  0
2014-09-01 09:36:00  6
8
                     A
2014-09-01 09:26:00  0
2014-09-01 09:34:00  4
16
                     A
2014-09-01 09:18:00  0
2014-09-01 09:34:00  4

[dimbeto]

Note that this is not the same as issue #8371 . Resample seems to work fine with any frequency but a multiple of 8.

[jreback]

@dimbeto not excatly the same as #8371, but its coming from the same code block. fix upcoming.

[jreback]

@dimbeto

if your first date were 09:30:01 how is pandas to know that you DON't want this
e.g. resample snaps the first date to the interval, so both get mapped to 09:28

Since the other intervals 'happen' to evenly divide your first date you are not noticing this.

their are several options to 'fix' this,

e.g. loffset=pd.Timedelta('2min') or base=2
will do what you want.

However, not sure that this is possible with the current option set to do automatically.

You almost want an option to base='start' which will use your own starting date if its an evently divisible snap interval.

[dimbeto]

Thanks.

Base='start' does look like a good idea. Are you assuming that the new frequency is a multiple of the old one?

And currently does pandas have a way to check that the index of a Series is, a concatenated sequence of, say 1 minute dateranges all starting from 9:30 on each day (but not necessarily ending at 16:00 for the case of holidays)? Currently, in order to create 7 minute bars from 1 minute bars, I would have to either create my own concatenated sequence of sub-dateranges, or I could group by day and use resample + between_time.

[jreback]

I will think about this for 0.15.1. But I think base='start' could align the resample to the first value so that the base freq (e.g. 1min) is the same as the first value. w/o having to specify base=2 in this case.

if you check s.index.freq if it is not None then it is a 'regular' freq. However gaps will cause this issues.

Their are various ways do this this, most notably:

In [33]: s = Series(np.arange(10),pd.date_range('20130101 09:30',periods=9,freq='1T').union([Timestamp('20130101 09:50')]))

In [34]: s
Out[34]: 
2013-01-01 09:30:00    0
2013-01-01 09:31:00    1
2013-01-01 09:32:00    2
2013-01-01 09:33:00    3
2013-01-01 09:34:00    4
2013-01-01 09:35:00    5
2013-01-01 09:36:00    6
2013-01-01 09:37:00    7
2013-01-01 09:38:00    8
2013-01-01 09:50:00    9
dtype: int64

In [35]: si = s.index.to_series()

In [36]: si
Out[36]: 
2013-01-01 09:30:00   2013-01-01 09:30:00
2013-01-01 09:31:00   2013-01-01 09:31:00
2013-01-01 09:32:00   2013-01-01 09:32:00
2013-01-01 09:33:00   2013-01-01 09:33:00
2013-01-01 09:34:00   2013-01-01 09:34:00
2013-01-01 09:35:00   2013-01-01 09:35:00
2013-01-01 09:36:00   2013-01-01 09:36:00
2013-01-01 09:37:00   2013-01-01 09:37:00
2013-01-01 09:38:00   2013-01-01 09:38:00
2013-01-01 09:50:00   2013-01-01 09:50:00
dtype: datetime64[ns]

In [37]: si-si.shift(1)
Out[37]: 
2013-01-01 09:30:00        NaT
2013-01-01 09:31:00   00:01:00
2013-01-01 09:32:00   00:01:00
2013-01-01 09:33:00   00:01:00
2013-01-01 09:34:00   00:01:00
2013-01-01 09:35:00   00:01:00
2013-01-01 09:36:00   00:01:00
2013-01-01 09:37:00   00:01:00
2013-01-01 09:38:00   00:01:00
2013-01-01 09:50:00   00:12:00
dtype: timedelta64[ns]

In [38]: (si-si.shift(1))==pd.Timedelta('1min')
Out[38]: 
2013-01-01 09:30:00    False
2013-01-01 09:31:00     True
2013-01-01 09:32:00     True
2013-01-01 09:33:00     True
2013-01-01 09:34:00     True
2013-01-01 09:35:00     True
2013-01-01 09:36:00     True
2013-01-01 09:37:00     True
2013-01-01 09:38:00     True
2013-01-01 09:50:00    False
dtype: bool

[dimbeto]

That works! Thanks.

[jreback]

@dimbeto great!