Rolling window and rolling apply with flexible lookback and resample behavior that also support parallelism.

custom_rolling.py

import pandas as pd

from pandas.tseries.offsets import *

def rolling_window(data, resample='1BM', lookback=12 * BMonthEnd()):
    dts = data.resample(resample).index
    if lookback == 'aggregate':
        for dt in dts:
            yield data.loc[:dt]
    else:
        dts = dts[dts > data.index[0] + lookback]
        for dt in dts:
            yield data.loc[dt-lookback:dt]
        
def rolling_apply(func, data, resample='1BM', lookback=12 * BMonthEnd(), column_wise=False, 
                  view=None, args=(), kwargs={}):
    if column_wise:
        return pd.concat([rolling_apply(func, data[col], resample=resample, 
                                        lookback=lookback, column_wise=False)
                          for col in data
                         ])
    else:
        if view is None:
            return pd.concat({data_window.index[-1]: func(data_window, *args, **kwargs)
                              for data_window in rolling_window(data, resample=resample, lookback=lookback)
                             }, axis=1).T
        else:
            workers = {}
            for data_window in rolling_window(data, resample=resample, lookback=lookback):
                workers[data_window.index[-1]] = view.apply_async(func, data_window, *args, **kwargs)
            
            return pd.concat({dt: worker.get() for dt, worker in workers.iteritems()}, axis=1).T