pandas rolling apply doesn't do anything

Question

I have a DataFrame like this:

df2 = pd.DataFrame({'date': ['2015-01-01', '2015-01-02', '2015-01-03'],
                    'value': ['a', 'b', 'a']})

         date value
0  2015-01-01     a
1  2015-01-02     b
2  2015-01-03     a

I'm trying to understand how to apply a custom rolling function to it. I've tried doing this:

df2.rolling(2).apply(lambda x: 1)

But this gives me the original DataFrame back:

         date value
0  2015-01-01     a
1  2015-01-02     b
2  2015-01-03     a

If I have a different DataFrame, like this:

df3 = pd.DataFrame({'a': [1, 2, 3], 'value': [4, 5, 6]})

The same rolling apply seems to work:

df3.rolling(2).apply(lambda x: 1)

     a  value
0  NaN    NaN
1  1.0    1.0
2  1.0    1.0

Why is this not working for the first DataFrame?

Pandas version: 0.20.2

Python version: 2.7.10

Update

So, I've realized that df2's columns are object-type, whereas the output of my lambda function is an integer. df3's columns are both integer columns. I'm assuming that this is why the apply isn't working.

The following doesn't work:

df2.rolling(2).apply(lambda x: 'a')
         date value
0  2015-01-01     a
1  2015-01-02     b
2  2015-01-03     a

Furthermore, say I want to concatenate the characters in the value column on a rolling basis, so that the output of the lambda function is a string, rather than an integer. The following also doesn't work:

df2.rolling(2).apply(lambda x: '.'.join(x))

         date value
0  2015-01-01     a
1  2015-01-02     b
2  2015-01-03     a

What's going on here? Can rolling operations be applied to object-type columns in pandas?

Answer 1

Here is one way this could be approached. Noting that rolling is a wrapper for numpy methods and the efficiency associated with those, this is not that. This merely provides a similiar api, to allow rolling on non-numeric columns:

Code:

import pandas as pd

class MyDataFrame(pd.DataFrame):

    @property
    def _constructor(self):
        return MyDataFrame

    def rolling_object(self, window, column, default):
        return pd.concat(
            [self[column].shift(i) for i in range(window)],
            axis=1).fillna(default).T

This creates a custom dataframe class that has a rolling_object method. It does not well match the pandas way in that it only operates on a single column at a time.

Test Code:

df2 = MyDataFrame({'date': ['2015-01-01', '2015-01-02', '2015-01-03'],
                   'value': ['a', 'b', 'c'],
                   'num': [1, 2, 3]
                   })

print(df2.rolling_object(2, 'value', '').apply(lambda x: '.'.join(x)))

Results:

0     a.
1    b.a
2    c.b
dtype: object