Extra Bin with Pandas Resample in 0.11.0

[waltaskew]

I've got a pandas data frame defined like this, using pandas 0.11.0:

    last_4_weeks_range = pandas.date_range(                                
            start=datetime.datetime(2001, 5, 4), periods=28)               
    last_4_weeks = pandas.DataFrame(                                       
        [{'REST_KEY': 1, 'DLY_TRN_QT': 80, 'DLY_SLS_AMT': 90,              
            'COOP_DLY_TRN_QT': 30, 'COOP_DLY_SLS_AMT': 20}] * 28 +         
        [{'REST_KEY': 2, 'DLY_TRN_QT': 70, 'DLY_SLS_AMT': 10,              
            'COOP_DLY_TRN_QT': 50, 'COOP_DLY_SLS_AMT': 20}] * 28,          
        index=last_4_weeks_range.append(last_4_weeks_range))               
    last_4_weeks.sort(inplace=True)

and when I go to resample it:

In [265]: last_4_weeks.resample('7D', how='sum')
Out[265]: 
            COOP_DLY_SLS_AMT  COOP_DLY_TRN_QT  DLY_SLS_AMT  DLY_TRN_QT  REST_KEY
2001-05-04               280              560          700        1050        21
2001-05-11               280              560          700        1050        21
2001-05-18               280              560          700        1050        21
2001-05-25               280              560          700        1050        21
2001-06-01                 0                0            0           0         0

I end up with an extra empty bin I wouldn't expect to see -- 2001-06-01. I wouldn't expect that bin to be there, as my 28 days are evenly divisible into the 7 day resample I'm performing. I've tried messing around with the closed kwarg, but I can't escape that extra bin. This seems like a bug, and it messes up my mean calculations when I try to do

In [266]: last_4_weeks.groupby('REST_KEY').resample('7D', how='sum').mean(level=0)
Out[266]: 
          COOP_DLY_SLS_AMT  COOP_DLY_TRN_QT  DLY_SLS_AMT  DLY_TRN_QT  REST_KEY
REST_KEY                                                                      
1                      112              168          504         448       5.6
2                      112              280           56         392      11.2

as the numbers are being divided by 5 rather than 4. (I also wouldn't expect REST_KEY to show up in the aggregation columns as it's part of the groupby, but that's really a smaller problem.)

[waltaskew]

This is curiously not the case if I pass how='count' -- no extra bin is returned. This makes me suspect a bug:

In [8]: last_4_weeks.resample('7D', how='count')
Out[8]: 
2001-05-04  COOP_DLY_SLS_AMT    14
            COOP_DLY_TRN_QT     14
            DLY_SLS_AMT         14
            DLY_TRN_QT          14
            REST_KEY            14
2001-05-11  COOP_DLY_SLS_AMT    14
            COOP_DLY_TRN_QT     14
            DLY_SLS_AMT         14
            DLY_TRN_QT          14
            REST_KEY            14
2001-05-18  COOP_DLY_SLS_AMT    14
            COOP_DLY_TRN_QT     14
            DLY_SLS_AMT         14
            DLY_TRN_QT          14
            REST_KEY            14
2001-05-25  COOP_DLY_SLS_AMT    14
            COOP_DLY_TRN_QT     14
            DLY_SLS_AMT         14
            DLY_TRN_QT          14
            REST_KEY            14
dtype: int64

[cpcloud]

a somewhat related issue in master is that there's no longer zeros there, there's garbage values.

this is a bug in how python vs. cythonized methods work, for example passing a lambda works

In [5]: last_4_weeks.resample('7D',how=lambda x:mean(x))
Out[5]:
            COOP_DLY_SLS_AMT  COOP_DLY_TRN_QT  DLY_SLS_AMT  DLY_TRN_QT  \
2001-05-04                20               40           50          75
2001-05-11                20               40           50          75
2001-05-18                20               40           50          75
2001-05-25                20               40           50          75

            REST_KEY
2001-05-04       1.5
2001-05-11       1.5
2001-05-18       1.5
2001-05-25       1.5

[waltaskew]

This also seems to act differently with different resample frequencies. With a frequency of 'AS', how='sum' yields the correct answer while how=lambda x: numpy.sum(x) does not:

In [14]: last_4_weeks.resample('AS', how='mean')
Out[14]: 
            COOP_DLY_SLS_AMT  COOP_DLY_TRN_QT  DLY_SLS_AMT  DLY_TRN_QT  REST_KEY
2001-01-01                20               40           50          75       1.5

In [15]: last_4_weeks.resample('AS', how=lambda x: numpy.mean(x))
Out[15]: 
            COOP_DLY_SLS_AMT  COOP_DLY_TRN_QT  DLY_SLS_AMT  DLY_TRN_QT  REST_KEY
2001-01-01               NaN              NaN          NaN         NaN       NaN

In [16]: last_4_weeks.resample('AS', how='sum')
Out[16]: 
            COOP_DLY_SLS_AMT  COOP_DLY_TRN_QT  DLY_SLS_AMT  DLY_TRN_QT  REST_KEY
2001-01-01              1120             2240         2800        4200        84

In [17]: last_4_weeks.resample('AS', how=lambda x: numpy.sum(x))
Out[17]: 
            COOP_DLY_SLS_AMT  COOP_DLY_TRN_QT  DLY_SLS_AMT  DLY_TRN_QT  REST_KEY
2001-01-01                 0                0            0           0         0

[cpcloud]

your last example is an issue with NaN handling

[krapfn]

I have also been having issues with resample adding extra bins (also in 0.11.0), and just thought I'd add that I can also see it even when the number of bins is not evenly divisible:

>>> x = pandas.DataFrame(numpy.random.randn(9, 3), index=pandas.date_range('2000-1-1', periods=9))
>>> x
                   0         1         2
2000-01-01 -1.191405  0.645320  1.308088
2000-01-02  1.229103 -0.727613  0.488344
2000-01-03  0.885808  1.381995 -0.955914
2000-01-04 -1.013526 -0.225070 -0.163507
2000-01-05  0.670316 -0.828281 -0.233381
2000-01-06  1.357537  1.446020 -0.661463
2000-01-07  0.335799  0.952127  0.591679
2000-01-08 -0.083534  1.025077 -0.146682
2000-01-09 -1.338294  1.919551  0.446385
>>> x.resample('5D')
                   0         1              2
2000-01-01  0.116059  0.049270   8.872589e-02
2000-01-06  0.067877  1.335694   5.747979e-02
2000-01-11  0.591679  0.146682  3.952525e-322

I don't have any particular insight to add, but maybe this extra info will help...

[jreback]

@cpcloud 0.13 or push?

[cpcloud]

like to do 0.13 but got a lot on my plate already ... let me see if there's anything else i can push to 0.14 in favor of this

[jreback]

up2u

[jreback]

pushing for now...can always pull back!

[cpcloud]

Ok
On Oct 4, 2013 4:21 PM, "jreback" [email protected] wrote:

pushing for now...can always pull back!

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/4076#issuecomment-25727391
.