Pandas: rolling mean by time interval

Question:

I’ve got a bunch of polling data; I want to compute a Pandas rolling mean to get an estimate for each day based on a three-day window. According to this question, the rolling_* functions compute the window based on a specified number of values, and not a specific datetime range.

How do I implement this functionality?

Sample input data:

polls_subset.tail(20)
Out[185]: 
            favorable  unfavorable  other

enddate                                  
2012-10-25       0.48         0.49   0.03
2012-10-25       0.51         0.48   0.02
2012-10-27       0.51         0.47   0.02
2012-10-26       0.56         0.40   0.04
2012-10-28       0.48         0.49   0.04
2012-10-28       0.46         0.46   0.09
2012-10-28       0.48         0.49   0.03
2012-10-28       0.49         0.48   0.03
2012-10-30       0.53         0.45   0.02
2012-11-01       0.49         0.49   0.03
2012-11-01       0.47         0.47   0.05
2012-11-01       0.51         0.45   0.04
2012-11-03       0.49         0.45   0.06
2012-11-04       0.53         0.39   0.00
2012-11-04       0.47         0.44   0.08
2012-11-04       0.49         0.48   0.03
2012-11-04       0.52         0.46   0.01
2012-11-04       0.50         0.47   0.03
2012-11-05       0.51         0.46   0.02
2012-11-07       0.51         0.41   0.00

Output would have only one row for each date.

Asked By: Anov

||

Answers:

What about something like this:

First resample the data frame into 1D intervals. This takes the mean of the values for all duplicate days. Use the fill_method option to fill in missing date values. Next, pass the resampled frame into pd.rolling_mean with a window of 3 and min_periods=1 :

pd.rolling_mean(df.resample("1D", fill_method="ffill"), window=3, min_periods=1)

            favorable  unfavorable     other
enddate
2012-10-25   0.495000     0.485000  0.025000
2012-10-26   0.527500     0.442500  0.032500
2012-10-27   0.521667     0.451667  0.028333
2012-10-28   0.515833     0.450000  0.035833
2012-10-29   0.488333     0.476667  0.038333
2012-10-30   0.495000     0.470000  0.038333
2012-10-31   0.512500     0.460000  0.029167
2012-11-01   0.516667     0.456667  0.026667
2012-11-02   0.503333     0.463333  0.033333
2012-11-03   0.490000     0.463333  0.046667
2012-11-04   0.494000     0.456000  0.043333
2012-11-05   0.500667     0.452667  0.036667
2012-11-06   0.507333     0.456000  0.023333
2012-11-07   0.510000     0.443333  0.013333

UPDATE: As Ben points out in the comments, with pandas 0.18.0 the syntax has changed. With the new syntax this would be:

df.resample("1d").sum().fillna(0).rolling(window=3, min_periods=1).mean()
Answered By: Zelazny7

I just had the same question but with irregularly spaced datapoints. Resample is not really an option here. So I created my own function. Maybe it will be useful for others too:

from pandas import Series, DataFrame
import pandas as pd
from datetime import datetime, timedelta
import numpy as np

def rolling_mean(data, window, min_periods=1, center=False):
    ''' Function that computes a rolling mean

    Parameters
    ----------
    data : DataFrame or Series
           If a DataFrame is passed, the rolling_mean is computed for all columns.
    window : int or string
             If int is passed, window is the number of observations used for calculating 
             the statistic, as defined by the function pd.rolling_mean()
             If a string is passed, it must be a frequency string, e.g. '90S'. This is
             internally converted into a DateOffset object, representing the window size.
    min_periods : int
                  Minimum number of observations in window required to have a value.

    Returns
    -------
    Series or DataFrame, if more than one column    
    '''
    def f(x):
        '''Function to apply that actually computes the rolling mean'''
        if center == False:
            dslice = col[x-pd.datetools.to_offset(window).delta+timedelta(0,0,1):x]
                # adding a microsecond because when slicing with labels start and endpoint
                # are inclusive
        else:
            dslice = col[x-pd.datetools.to_offset(window).delta/2+timedelta(0,0,1):
                         x+pd.datetools.to_offset(window).delta/2]
        if dslice.size < min_periods:
            return np.nan
        else:
            return dslice.mean()

    data = DataFrame(data.copy())
    dfout = DataFrame()
    if isinstance(window, int):
        dfout = pd.rolling_mean(data, window, min_periods=min_periods, center=center)
    elif isinstance(window, basestring):
        idx = Series(data.index.to_pydatetime(), index=data.index)
        for colname, col in data.iterkv():
            result = idx.apply(f)
            result.name = colname
            dfout = dfout.join(result, how='outer')
    if dfout.columns.size == 1:
        dfout = dfout.ix[:,0]
    return dfout


# Example
idx = [datetime(2011, 2, 7, 0, 0),
       datetime(2011, 2, 7, 0, 1),
       datetime(2011, 2, 7, 0, 1, 30),
       datetime(2011, 2, 7, 0, 2),
       datetime(2011, 2, 7, 0, 4),
       datetime(2011, 2, 7, 0, 5),
       datetime(2011, 2, 7, 0, 5, 10),
       datetime(2011, 2, 7, 0, 6),
       datetime(2011, 2, 7, 0, 8),
       datetime(2011, 2, 7, 0, 9)]
idx = pd.Index(idx)
vals = np.arange(len(idx)).astype(float)
s = Series(vals, index=idx)
rm = rolling_mean(s, window='2min')
Answered By: user2689410

user2689410’s code was exactly what I needed. Providing my version (credits to user2689410), which is faster due to calculating mean at once for whole rows in the DataFrame.

Hope my suffix conventions are readable: _s: string, _i: int, _b: bool, _ser: Series and _df: DataFrame. Where you find multiple suffixes, type can be both.

import pandas as pd
from datetime import datetime, timedelta
import numpy as np

def time_offset_rolling_mean_df_ser(data_df_ser, window_i_s, min_periods_i=1, center_b=False):
    """ Function that computes a rolling mean

    Credit goes to user2689410 at http://stackoverflow.com/questions/15771472/pandas-rolling-mean-by-time-interval

    Parameters
    ----------
    data_df_ser : DataFrame or Series
         If a DataFrame is passed, the time_offset_rolling_mean_df_ser is computed for all columns.
    window_i_s : int or string
         If int is passed, window_i_s is the number of observations used for calculating
         the statistic, as defined by the function pd.time_offset_rolling_mean_df_ser()
         If a string is passed, it must be a frequency string, e.g. '90S'. This is
         internally converted into a DateOffset object, representing the window_i_s size.
    min_periods_i : int
         Minimum number of observations in window_i_s required to have a value.

    Returns
    -------
    Series or DataFrame, if more than one column

    >>> idx = [
    ...     datetime(2011, 2, 7, 0, 0),
    ...     datetime(2011, 2, 7, 0, 1),
    ...     datetime(2011, 2, 7, 0, 1, 30),
    ...     datetime(2011, 2, 7, 0, 2),
    ...     datetime(2011, 2, 7, 0, 4),
    ...     datetime(2011, 2, 7, 0, 5),
    ...     datetime(2011, 2, 7, 0, 5, 10),
    ...     datetime(2011, 2, 7, 0, 6),
    ...     datetime(2011, 2, 7, 0, 8),
    ...     datetime(2011, 2, 7, 0, 9)]
    >>> idx = pd.Index(idx)
    >>> vals = np.arange(len(idx)).astype(float)
    >>> ser = pd.Series(vals, index=idx)
    >>> df = pd.DataFrame({'s1':ser, 's2':ser+1})
    >>> time_offset_rolling_mean_df_ser(df, window_i_s='2min')
                          s1   s2
    2011-02-07 00:00:00  0.0  1.0
    2011-02-07 00:01:00  0.5  1.5
    2011-02-07 00:01:30  1.0  2.0
    2011-02-07 00:02:00  2.0  3.0
    2011-02-07 00:04:00  4.0  5.0
    2011-02-07 00:05:00  4.5  5.5
    2011-02-07 00:05:10  5.0  6.0
    2011-02-07 00:06:00  6.0  7.0
    2011-02-07 00:08:00  8.0  9.0
    2011-02-07 00:09:00  8.5  9.5
    """

    def calculate_mean_at_ts(ts):
        """Function (closure) to apply that actually computes the rolling mean"""
        if center_b == False:
            dslice_df_ser = data_df_ser[
                ts-pd.datetools.to_offset(window_i_s).delta+timedelta(0,0,1):
                ts
            ]
            # adding a microsecond because when slicing with labels start and endpoint
            # are inclusive
        else:
            dslice_df_ser = data_df_ser[
                ts-pd.datetools.to_offset(window_i_s).delta/2+timedelta(0,0,1):
                ts+pd.datetools.to_offset(window_i_s).delta/2
            ]
        if  (isinstance(dslice_df_ser, pd.DataFrame) and dslice_df_ser.shape[0] < min_periods_i) or 
            (isinstance(dslice_df_ser, pd.Series) and dslice_df_ser.size < min_periods_i):
            return dslice_df_ser.mean()*np.nan   # keeps number format and whether Series or DataFrame
        else:
            return dslice_df_ser.mean()

    if isinstance(window_i_s, int):
        mean_df_ser = pd.rolling_mean(data_df_ser, window=window_i_s, min_periods=min_periods_i, center=center_b)
    elif isinstance(window_i_s, basestring):
        idx_ser = pd.Series(data_df_ser.index.to_pydatetime(), index=data_df_ser.index)
        mean_df_ser = idx_ser.apply(calculate_mean_at_ts)

    return mean_df_ser
Answered By: Mark Horvath

I found that user2689410 code broke when I tried with window=’1M’ as the delta on business month threw this error:

AttributeError: 'MonthEnd' object has no attribute 'delta'

I added the option to pass directly a relative time delta, so you can do similar things for user defined periods.

Thanks for the pointers, here’s my attempt – hope it’s of use.

def rolling_mean(data, window, min_periods=1, center=False):
""" Function that computes a rolling mean
Reference:
    http://stackoverflow.com/questions/15771472/pandas-rolling-mean-by-time-interval

Parameters
----------
data : DataFrame or Series
       If a DataFrame is passed, the rolling_mean is computed for all columns.
window : int, string, Timedelta or Relativedelta
         int - number of observations used for calculating the statistic,
               as defined by the function pd.rolling_mean()
         string - must be a frequency string, e.g. '90S'. This is
                  internally converted into a DateOffset object, and then
                  Timedelta representing the window size.
         Timedelta / Relativedelta - Can directly pass a timedeltas.
min_periods : int
              Minimum number of observations in window required to have a value.
center : bool
         Point around which to 'center' the slicing.

Returns
-------
Series or DataFrame, if more than one column
"""
def f(x, time_increment):
    """Function to apply that actually computes the rolling mean
    :param x:
    :return:
    """
    if not center:
        # adding a microsecond because when slicing with labels start
        # and endpoint are inclusive
        start_date = x - time_increment + timedelta(0, 0, 1)
        end_date = x
    else:
        start_date = x - time_increment/2 + timedelta(0, 0, 1)
        end_date = x + time_increment/2
    # Select the date index from the
    dslice = col[start_date:end_date]

    if dslice.size < min_periods:
        return np.nan
    else:
        return dslice.mean()

data = DataFrame(data.copy())
dfout = DataFrame()
if isinstance(window, int):
    dfout = pd.rolling_mean(data, window, min_periods=min_periods, center=center)

elif isinstance(window, basestring):
    time_delta = pd.datetools.to_offset(window).delta
    idx = Series(data.index.to_pydatetime(), index=data.index)
    for colname, col in data.iteritems():
        result = idx.apply(lambda x: f(x, time_delta))
        result.name = colname
        dfout = dfout.join(result, how='outer')

elif isinstance(window, (timedelta, relativedelta)):
    time_delta = window
    idx = Series(data.index.to_pydatetime(), index=data.index)
    for colname, col in data.iteritems():
        result = idx.apply(lambda x: f(x, time_delta))
        result.name = colname
        dfout = dfout.join(result, how='outer')

if dfout.columns.size == 1:
    dfout = dfout.ix[:, 0]
return dfout

And the example with a 3 day time window to calculate the mean:

from pandas import Series, DataFrame
import pandas as pd
from datetime import datetime, timedelta
import numpy as np
from dateutil.relativedelta import relativedelta

idx = [datetime(2011, 2, 7, 0, 0),
           datetime(2011, 2, 7, 0, 1),
           datetime(2011, 2, 8, 0, 1, 30),
           datetime(2011, 2, 9, 0, 2),
           datetime(2011, 2, 10, 0, 4),
           datetime(2011, 2, 11, 0, 5),
           datetime(2011, 2, 12, 0, 5, 10),
           datetime(2011, 2, 12, 0, 6),
           datetime(2011, 2, 13, 0, 8),
           datetime(2011, 2, 14, 0, 9)]
idx = pd.Index(idx)
vals = np.arange(len(idx)).astype(float)
s = Series(vals, index=idx)
# Now try by passing the 3 days as a relative time delta directly.
rm = rolling_mean(s, window=relativedelta(days=3))
>>> rm
Out[2]: 
2011-02-07 00:00:00    0.0
2011-02-07 00:01:00    0.5
2011-02-08 00:01:30    1.0
2011-02-09 00:02:00    1.5
2011-02-10 00:04:00    3.0
2011-02-11 00:05:00    4.0
2011-02-12 00:05:10    5.0
2011-02-12 00:06:00    5.5
2011-02-13 00:08:00    6.5
2011-02-14 00:09:00    7.5
Name: 0, dtype: float64
Answered By: InterwebIsGreat

This example seems to call for a weighted mean as suggested in @andyhayden’s comment. For example, there are two polls on 10/25 and one each on 10/26 and 10/27. If you just resample and then take the mean, this effectively gives twice as much weighting to the polls on 10/26 and 10/27 compared to the ones on 10/25.

To give equal weight to each poll rather than equal weight to each day, you could do something like the following.

>>> wt = df.resample('D',limit=5).count()

            favorable  unfavorable  other
enddate                                  
2012-10-25          2            2      2
2012-10-26          1            1      1
2012-10-27          1            1      1

>>> df2 = df.resample('D').mean()

            favorable  unfavorable  other
enddate                                  
2012-10-25      0.495        0.485  0.025
2012-10-26      0.560        0.400  0.040
2012-10-27      0.510        0.470  0.020

That gives you the raw ingredients for doing a poll-based mean instead of a day-based mean. As before, the polls are averaged on 10/25, but the weight for 10/25 is also stored and is double the weight on 10/26 or 10/27 to reflect that two polls were taken on 10/25.

>>> df3 = df2 * wt
>>> df3 = df3.rolling(3,min_periods=1).sum()
>>> wt3 = wt.rolling(3,min_periods=1).sum()

>>> df3 = df3 / wt3  

            favorable  unfavorable     other
enddate                                     
2012-10-25   0.495000     0.485000  0.025000
2012-10-26   0.516667     0.456667  0.030000
2012-10-27   0.515000     0.460000  0.027500
2012-10-28   0.496667     0.465000  0.041667
2012-10-29   0.484000     0.478000  0.042000
2012-10-30   0.488000     0.474000  0.042000
2012-10-31   0.530000     0.450000  0.020000
2012-11-01   0.500000     0.465000  0.035000
2012-11-02   0.490000     0.470000  0.040000
2012-11-03   0.490000     0.465000  0.045000
2012-11-04   0.500000     0.448333  0.035000
2012-11-05   0.501429     0.450000  0.032857
2012-11-06   0.503333     0.450000  0.028333
2012-11-07   0.510000     0.435000  0.010000

Note that the rolling mean for 10/27 is now 0.51500 (poll-weighted) rather than 52.1667 (day-weighted).

Also note that there have been changes to the APIs for resample and rolling as of version 0.18.0.

rolling (what’s new in pandas 0.18.0)

resample (what’s new in pandas 0.18.0)

Answered By: JohnE

In the meantime, a time-window capability was added. See this link.

In [1]: df = DataFrame({'B': range(5)})

In [2]: df.index = [Timestamp('20130101 09:00:00'),
   ...:             Timestamp('20130101 09:00:02'),
   ...:             Timestamp('20130101 09:00:03'),
   ...:             Timestamp('20130101 09:00:05'),
   ...:             Timestamp('20130101 09:00:06')]

In [3]: df
Out[3]: 
                     B
2013-01-01 09:00:00  0
2013-01-01 09:00:02  1
2013-01-01 09:00:03  2
2013-01-01 09:00:05  3
2013-01-01 09:00:06  4

In [4]: df.rolling(2, min_periods=1).sum()
Out[4]: 
                       B
2013-01-01 09:00:00  0.0
2013-01-01 09:00:02  1.0
2013-01-01 09:00:03  3.0
2013-01-01 09:00:05  5.0
2013-01-01 09:00:06  7.0

In [5]: df.rolling('2s', min_periods=1).sum()
Out[5]: 
                       B
2013-01-01 09:00:00  0.0
2013-01-01 09:00:02  1.0
2013-01-01 09:00:03  3.0
2013-01-01 09:00:05  3.0
2013-01-01 09:00:06  7.0
Answered By: Martin

To keep it basic, I used a loop and something like this to get you started (my index are datetimes):

import pandas as pd
import datetime as dt

#populate your dataframe: "df"
#...

df[df.index<(df.index[0]+dt.timedelta(hours=1))] #gives you a slice. you can then take .sum() .mean(), whatever

and then you can run functions on that slice. You can see how adding an iterator to make the start of the window something other than the first value in your dataframes index would then roll the window (you could use a > rule for the start as well for example).

Note, this may be less efficient for SUPER large data or very small increments as your slicing may become more strenuous (works for me well enough for hundreds of thousands of rows of data and several columns though for hourly windows across a few weeks)

Answered By: Vlox

Check that your index is really datetime, not str
Can be helpful:

data.index = pd.to_datetime(data['Index']).values
Answered By: evgps

visualize the rolling averages to see if it makes sense. I don’t understand why sum was used when the rolling average was requested.

  df=pd.read_csv('poll.csv',parse_dates=['enddate'],dtype={'favorable':np.float,'unfavorable':np.float,'other':np.float})

  df.set_index('enddate')
  df=df.fillna(0)

 fig, axs = plt.subplots(figsize=(5,10))
 df.plot(x='enddate', ax=axs)
 plt.show()


 df.rolling(window=3,min_periods=3).mean().plot()
 plt.show()
 print("The larger the window coefficient the smoother the line will appear")
 print('The min_periods is the minimum number of observations in the window required to have a value')

 df.rolling(window=6,min_periods=3).mean().plot()
 plt.show()
Answered By: Golden Lion