BUG: Unexpected period range index shift

[jmcinto]

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the master branch of pandas.

Reproducible Example

import pandas as pd
from datetime import datetime
print(pd.__version__)
df = pd.DataFrame(range(24))
df.index = pd.period_range(datetime(1900,1,1), periods=24, freq="1H")
resamp = df.resample("6H",label='right', closed='right').sum()

Issue Description

Resampling a df that has a period range index results in an unexpected temporal shift in the results in the pandas version 1.3.4. This can be seen by running the example code in version 1.3.4 and comparing the results to version 1.0.5.

Expected Behavior

Running the example code in version 1.0.5 results in the expected behavior.

Installed Versions

1.3.4

[mroeschke]

Could you give this issue a more informative title and post a code output from 1.0.5?

[jmcinto]

@mroeschke see below for a better overview of results of running sample code. Apologies for neglecting to add a useful title!

Sample Code (slight modification to clarify result being compared):

import pandas as pd
from datetime import datetime
print(pd.version)
df = pd.DataFrame(range(24))
df.index = pd.period_range(datetime(1900,1,1), periods=24, freq="1H")
resamp = df.resample("6H",label='right', closed='right').sum()
print('resamp:\n',resamp)

Results when running with 1.0.5:
1.0.5
resamp:
0
1900-01-01 00:00 15
1900-01-01 06:00 51
1900-01-01 12:00 87
1900-01-01 18:00 123

Results when running in 1.3.4:
1.3.4
resamp:
0
1899-12-31 18:00 15
1900-01-01 00:00 51
1900-01-01 06:00 87
1900-01-01 12:00 123
1900-01-01 18:00 0

[sappersapper]

My test shows the same difference between v1.0.5 and v1.2.0. It seems either setting label='right' or closed='right' will lead to different results between v1.0.5 and v1.2.0.

It's bothering me how to understand closed='right' for PeriodIndex downsampling. It seems not well defined.

Try to make the it more clear:

def print_series(series):
    print(pd.DataFrame({'period_start': series.index.start_time,
                        'period_end': series.index.end_time,
                        'value': series.values
                       },
                       index=series.index))

# The data for resampling
series = pd.Series(range(1, 5), index=pd.period_range('1/1/2000', periods=4, freq='1H'))
print_series(series)

	period_start	period_end	value
2000-01-01 00:00	2000-01-01 00:00:00	2000-01-01 00:59:59.999999999	1
2000-01-01 01:00	2000-01-01 01:00:00	2000-01-01 01:59:59.999999999	2
2000-01-01 02:00	2000-01-01 02:00:00	2000-01-01 02:59:59.999999999	3
2000-01-01 03:00	2000-01-01 03:00:00	2000-01-01 03:59:59.999999999	4

The default parameters closed='left' lead to same result.

print_series(series.resample('2H', closed='left').sum())

	period_start	period_end	value
2000-01-01 00:00	2000-01-01 00:00:00	2000-01-01 01:59:59.999999999	3
2000-01-01 02:00	2000-01-01 02:00:00	2000-01-01 03:59:59.999999999	7

While for closed='right', v1.0.5 and v1.2.0 behave differently:

print_series(series.resample('2H', closed='right').sum())

output of v1.0.5:

	period_start	period_end	value
2000-01-01 00:00	2000-01-01 00:00:00	2000-01-01 01:59:59.999999999	3
2000-01-01 02:00	2000-01-01 02:00:00	2000-01-01 03:59:59.999999999	7

output of v1.2.0:

	period_start	period_end	value
1999-12-31 22:00	1999-12-31 22:00:00	1999-12-31 23:59:59.999999999	3
2000-01-01 00:00	2000-01-01 00:00:00	2000-01-01 01:59:59.999999999	7
2000-01-01 02:00	2000-01-01 02:00:00	2000-01-01 03:59:59.999999999	0

Note the PeriodIndex is a left closed right open interval.

The output of v1.2.0 seem not resonable: all the input data are at 2000, while the resample from 1999-12-31 22:00:00 to 1999-12-31 23:59:59 have the sum 3.

As for v1.0.5, the result of closed='right' have no difference with result of closed='left'. This behavior seems not compatible with DatetimeIndex of closed='right'.

Maybe this is a more resonable ouput for print_series(series.resample('2H', closed='right').sum()) ?

	period_start	period_end	value
1999-12-31 23:00	1999-12-31 23:00:00	2000-01-01 00:59:59.999999999	1
2000-01-01 01:00	2000-01-01 01:00:00	2000-01-01 02:59:59.999999999	5
2000-01-01 03:00	2000-01-01 03:00:00	2000-01-01 04:59:59.999999999	4