Pandas DataFrame with MultiIndex: Group by year of DateTime level values

Question

I have and pandas dataframe with a multiindex that looks like this:

# -*- coding: utf-8 -*-
import numpy as np
import pandas as pd

# multi-indexed dataframe
df = pd.DataFrame(np.random.randn(8760 * 3, 3))
df['concept'] = "some_value"
df['datetime'] = pd.date_range(start='2016', periods=len(df), freq='60Min')
df.set_index(['concept', 'datetime'], inplace=True)
df.sort_index(inplace=True)

Console output:

df.head()
Out[23]: 
                 0         1         2
datetime                              
2016      0.458802  0.413004  0.091056
2016     -0.051840 -1.780310 -0.304122
2016     -1.119973  0.954591  0.279049
2016     -0.691850 -0.489335  0.554272
2016     -1.278834 -1.292012 -0.637931

df.head()
    ...: df.tail()

Out[24]: 
                 0         1         2
datetime                              
2018     -1.872155  0.434520 -0.526520
2018      0.345213  0.989475 -0.892028
2018     -0.162491  0.908121 -0.993499
2018     -1.094727  0.307312  0.515041
2018     -0.880608 -1.065203 -1.438645

Now I want to create annual sums along the level 'datetime'.

My first try was the following but this doesn't work:

# sum along years
years = df.index.get_level_values('datetime').year.tolist()
df.index.set_levels([years], level=['datetime'], inplace=True)
df = df.groupby(level=['datetime']).sum()

And it also seems quite heavy handed to me as this task is probably pretty easy to realize.

So here's my question: How can I get annual sums for the level 'datetime'? Is there a simple way to realize this by applying a function to the DateTime level values?

Answer 1

You can groupby by second level of multiindex and year:

# -*- coding: utf-8 -*-
import numpy as np
import pandas as pd

# multi-indexed dataframe
df = pd.DataFrame(np.random.randn(8760  * 3, 3))
df['concept'] = "some_value"
df['datetime'] = pd.date_range(start='2016', periods=len(df), freq='60Min')
df.set_index(['concept', 'datetime'], inplace=True)
df.sort_index(inplace=True)
print df.head() 
                                       0         1         2
concept    datetime                                         
some_value 2016-01-01 00:00:00  1.973437  0.101535 -0.693360
           2016-01-01 01:00:00  1.221657 -1.983806 -0.075609
           2016-01-01 02:00:00 -0.208122 -2.203801  1.254084
           2016-01-01 03:00:00  0.694332 -0.235864  0.538468
           2016-01-01 04:00:00 -0.928815 -1.417445  1.534218

# sum along years
#years = df.index.get_level_values('datetime').year.tolist()
#df.index.set_levels([years], level=['datetime'], inplace=True)

print df.index.levels[1].year
[2016 2016 2016 ..., 2018 2018 2018]
df = df.groupby(df.index.levels[1].year).sum()
print df.head()
               0           1          2
2016  -93.901914  -32.205514 -22.460965
2017  205.681817   67.701669 -33.960801
2018   67.438355  150.954614 -21.381809

Or you can use get_level_values and year:

df = df.groupby(df.index.get_level_values('datetime').year).sum()
print df.head()
               0           1          2
2016  -93.901914  -32.205514 -22.460965
2017  205.681817   67.701669 -33.960801
2018   67.438355  150.954614 -21.381809

Answer 2

Starting with your sample data:

df = pd.DataFrame(np.random.randn(8760 * 3, 3))
df['concept'] = "some_value"
df['datetime'] = pd.date_range(start='2016', periods=len(df), freq='60Min')
df.set_index(['concept', 'datetime'], inplace=True)

you can apply groupby to a level of your MultiIndex:

df.groupby(pd.TimeGrouper(level='datetime', freq='A')).sum()

to get:

                     0          1          2
datetime                                    
2016-12-31  100.346135 -71.673222  42.816675
2017-12-31 -132.880909 -66.017010 -73.449358
2018-12-31  -71.449710 -15.774929  97.634349

pd.TimeGrouper is now (0.23) deprecated; please use pd.Grouper(freq=...) instead.