Can I resample (ffill) pandas DataFrame with MultiIndex

Question

I have a dataframe that looks like this, with a MultiIndex over ('timestamp', 'id') :

                 value
timestamp   id
2020-03-03  A    100
2020-03-03  B    222
2020-03-03  C    5000
2020-03-04  A    NaN
2020-03-04  B    1
2020-03-04  C    NaN
2020-03-05  A    200
2020-03-05  B    NaN
2020-03-05  C    NaN
2020-03-06  A    NaN
2020-03-06  B    20
2020-03-06  C    NaN

I want to forwards fill (timewise) on value so that the dataframe is populated with the most recently available data item, ie the DataFrame becomes:

                 value
timestamp   id
2020-03-03  A    100
2020-03-03  B    222
2020-03-03  C    5000
2020-03-04  A    100
2020-03-04  B    1
2020-03-04  C    5000
2020-03-05  A    200
2020-03-05  B    1
2020-03-05  C    5000
2020-03-06  A    200
2020-03-06  B    20
2020-03-06  C    5000

Is there any easy way using resampler?

Answer 1

You can sort the second level and ffill , then reindex like original:

df.sort_index(level=1).ffill().reindex(df.index)

---

                value
timestamp  id        
2020-03-03 A    100.0
           B    222.0
           C   5000.0
2020-03-04 A    100.0
           B      1.0
           C   5000.0
2020-03-05 A    200.0
           B      1.0
           C   5000.0
2020-03-06 A    200.0
           B     20.0
           C   5000.0

Answer 2

You can also use stack to arrange the data in a correct 2D representation for filling (column-wise) and then unstack back to the original format. This treats columns (ie indexes) separately as opposed to rolling over data values, which is the case in the other solution given.

a = ['2020-03-03','2020-03-04','2020-03-05', '2020-03-06']
b = ['A', 'B', 'C']
c = ['value1', 'value2']
df = pd.DataFrame(data=None, index=pd.MultiIndex.from_product([a,b]), columns=c)
df.loc[('2020-03-03', slice(None)), 'value1'] = np.array([100, 222, 5000])
df.loc[('2020-03-04', 'B'), 'value1'] = 1.0
df.loc[('2020-03-05', 'A'), 'value1'] = 200.0
df.loc[('2020-03-06', 'C'), 'value1'] = 20
df['value2'] = df['value1']
df.loc[('2020-03-03', 'C'), 'value2'] = np.nan
df

                 value1  value2
timestamp   id
2020-03-03  A    100     100
2020-03-03  B    222     222
2020-03-03  C    5000    NaN   # <- OBS!
2020-03-04  A    NaN     NaN
2020-03-04  B    1       1
2020-03-04  C    NaN     NaN
2020-03-05  A    200     200
2020-03-05  B    NaN     NaN
2020-03-05  C    NaN     NaN
2020-03-06  A    NaN     NaN
2020-03-06  B    20      20
2020-03-06  C    NaN     NaN

Using df.unstack().fillna(method='ffill') gives

            value1             value2
            A     B     C      A     B     C
timestamp
2020-03-03  100   222  5000    100   222   NaN
2020-03-04  100   1    5000    100   1     NaN
2020-03-05  200   1    5000    200   1     NaN
2020-03-06  200   1    20      200   1     20

This can be reverted with .stack() to the original format again.

Comparing this to df.sort_index(level=1).ffill().reindex(df.index) the difference is in the last column where since 'C' start with an NaN the value from 'B' of 1 is rolled into the start of 'C' for 'Value2'.