Windowing Operations

pandas contains a compact set of APIs for performing windowing operations - an operation that performs an aggregation over a sliding partition of values. The API functions similarly to the groupby API in that Series and DataFrame call the windowing method with necessary parameters and then subsequently call the aggregation function.

In [1]: s = pd.Series(range(5))

In [2]: s.rolling(window=2).sum()
Out[2]: 
0    NaN
1    1.0
2    3.0
3    5.0
4    7.0
dtype: float64

The windows are comprised by looking back the length of the window from the current observation. The result above can be derived by taking the sum of the following windowed partitions of data:

In [3]: for window in s.rolling(window=2):
   ...:     print(window)
   ...: 
0    0
dtype: int64
0    0
1    1
dtype: int64
1    1
2    2
dtype: int64
2    2
3    3
dtype: int64
3    3
4    4
dtype: int64

Overview

pandas supports 4 types of windowing operations:

1. Rolling window: Generic fixed or variable sliding window over the values.

2. Weighted window: Weighted, non-rectangular window supplied by the scipy.signal library.

3. Expanding window: Accumulating window over the values.

4. Exponentially Weighted window: Accumulating and exponentially weighted window over the values.

Concept	Method	Returned Object	Supports time-based windows	Supports chained groupby
Rolling window	rolling	Rolling	Yes	Yes
Weighted window	rolling	Window	No	No
Expanding window	expanding	Expanding	No	Yes
Exponentially Weighted window	ewm	ExponentialMovingWindow	No	Yes (as of version 1.2)

As noted above, some operations support specifying a window based on a time offset:

In [4]: s = pd.Series(range(5), index=pd.date_range('2020-01-01', periods=5, freq='1D'))

In [5]: s.rolling(window='2D').sum()
Out[5]: 
2020-01-01    0.0
2020-01-02    1.0
2020-01-03    3.0
2020-01-04    5.0
2020-01-05    7.0
Freq: D, dtype: float64

Additionally, some methods support chaining a groupby operation with a windowing operation which will first group the data by the specified keys and then perform a windowing operation per group.

In [6]: df = pd.DataFrame({'A': ['a', 'b', 'a', 'b', 'a'], 'B': range(5)})

In [7]: df.groupby('A').expanding().sum()
Out[7]: 
       B
A       
a 0  0.0
  2  2.0
  4  6.0
b 1  1.0
  3  4.0

Note

Windowing operations currently only support numeric data (integer and float) and will always return float64 values.

Warning

Some windowing aggregation, mean, sum, var and std methods may suffer from numerical imprecision due to the underlying windowing algorithms accumulating sums. When values differ with magnitude \(1/np.finfo(np.double).eps\) this results in truncation. It must be noted, that large values may have an impact on windows, which do not include these values. Kahan summation is used to compute the rolling sums to preserve accuracy as much as possible.

All windowing operations support a min_periods argument that dictates the minimum amount of non-np.nan values a window must have; otherwise, the resulting value is np.nan. min_peridos defaults to 1 for time-based windows and window for fixed windows

In [8]: s = pd.Series([np.nan, 1, 2, np.nan, np.nan, 3])

In [9]: s.rolling(window=3, min_periods=1).sum()
Out[9]: 
0    NaN
1    1.0
2    3.0
3    3.0
4    2.0
5    3.0
dtype: float64

In [10]: s.rolling(window=3, min_periods=2).sum()
Out[10]: 
0    NaN
1    NaN
2    3.0
3    3.0
4    NaN
5    NaN
dtype: float64

# Equivalent to min_periods=3
In [11]: s.rolling(window=3, min_periods=None).sum()
Out[11]: 
0   NaN
1   NaN
2   NaN
3   NaN
4   NaN
5   NaN
dtype: float64

Additionally, all windowing operations supports the aggregate method for returning a result of multiple aggregations applied to a window.

In [12]: df = pd.DataFrame({"A": range(5), "B": range(10, 15)})

In [13]: df.expanding().agg([np.sum, np.mean, np.std])
Out[13]: 
      A                    B                
    sum mean       std   sum  mean       std
0   0.0  0.0       NaN  10.0  10.0       NaN
1   1.0  0.5  0.707107  21.0  10.5  0.707107
2   3.0  1.0  1.000000  33.0  11.0  1.000000
3   6.0  1.5  1.290994  46.0  11.5  1.290994
4  10.0  2.0  1.581139  60.0  12.0  1.581139

Rolling window

Generic rolling windows support specifying windows as a fixed number of observations or variable number of observations based on an offset. If a time based offset is provided, the corresponding time based index must be monotonic.

In [14]: times = ['2020-01-01', '2020-01-03', '2020-01-04', '2020-01-05', '2020-01-29']

In [15]: s = pd.Series(range(5), index=pd.DatetimeIndex(times))

In [16]: s
Out[16]: 
2020-01-01    0
2020-01-03    1
2020-01-04    2
2020-01-05    3
2020-01-29    4
dtype: int64

# Window with 2 observations
In [17]: s.rolling(window=2).sum()
Out[17]: 
2020-01-01    NaN
2020-01-03    1.0
2020-01-04    3.0
2020-01-05    5.0
2020-01-29    7.0
dtype: float64

# Window with 2 days worth of observations
In [18]: s.rolling(window='2D').sum()
Out[18]: 
2020-01-01    0.0
2020-01-03    1.0
2020-01-04    3.0
2020-01-05    5.0
2020-01-29    4.0
dtype: float64

For all supported aggregation functions, see Rolling window functions.

Centering windows

By default the labels are set to the right edge of the window, but a center keyword is available so the labels can be set at the center.

In [19]: s = pd.Series(range(10))

In [20]: s.rolling(window=5).mean()
Out[20]: 
0    NaN
1    NaN
2    NaN
3    NaN
4    2.0
5    3.0
6    4.0
7    5.0
8    6.0
9    7.0
dtype: float64

In [21]: s.rolling(window=5, center=True).mean()
Out[21]: 
0    NaN
1    NaN
2    2.0
3    3.0
4    4.0
5    5.0
6    6.0
7    7.0
8    NaN
9    NaN
dtype: float64

Rolling window endpoints

The inclusion of the interval endpoints in rolling window calculations can be specified with the closed parameter:

Value	Behavior
right'	close right endpoint
'left'	close left endpoint
'both'	close both endpoints
'neither'	open endpoints

For example, having the right endpoint open is useful in many problems that require that there is no contamination from present information back to past information. This allows the rolling window to compute statistics “up to that point in time”, but not including that point in time.

In [22]: df = pd.DataFrame(
   ....:     {"x": 1},
   ....:     index=[
   ....:         pd.Timestamp("20130101 09:00:01"),
   ....:         pd.Timestamp("20130101 09:00:02"),
   ....:         pd.Timestamp("20130101 09:00:03"),
   ....:         pd.Timestamp("20130101 09:00:04"),
   ....:         pd.Timestamp("20130101 09:00:06"),
   ....:     ],
   ....: )
   ....: 

In [23]: df["right"] = df.rolling("2s", closed="right").x.sum()  # default

In [24]: df["both"] = df.rolling("2s", closed="both").x.sum()

In [25]: df["left"] = df.rolling("2s", closed="left").x.sum()

In [26]: df["neither"] = df.rolling("2s", closed="neither").x.sum()

In [27]: df
Out[27]: 
                     x  right  both  left  neither
2013-01-01 09:00:01  1    1.0   1.0   NaN      NaN
2013-01-01 09:00:02  1    2.0   2.0   1.0      1.0
2013-01-01 09:00:03  1    2.0   3.0   2.0      1.0
2013-01-01 09:00:04  1    2.0   3.0   2.0      1.0
2013-01-01 09:00:06  1    1.0   2.0   1.0      NaN

Custom window rolling

New in version 1.0.

In addition to accepting an integer or offset as a window argument, rolling also accepts a BaseIndexer subclass that allows a user to define a custom method for calculating window bounds. The BaseIndexer subclass will need to define a get_window_bounds method that returns a tuple of two arrays, the first being the starting indices of the windows and second being the ending indices of the windows. Additionally, num_values, min_periods, center, closed and will automatically be passed to get_window_bounds and the defined method must always accept these arguments.

For example, if we have the following :class:DataFrame:

In [28]: use_expanding = [True, False, True, False, True]

In [29]: use_expanding
Out[29]: [True, False, True, False, True]

In [30]: df = pd.DataFrame({"values": range(5)})

In [31]: df
Out[31]: 
   values
0       0
1       1
2       2
3       3
4       4

and we want to use an expanding window where use_expanding is True otherwise a window of size 1, we can create the following BaseIndexer subclass:

In [2]: from pandas.api.indexers import BaseIndexer
...:
...: class CustomIndexer(BaseIndexer):
...:
...:    def get_window_bounds(self, num_values, min_periods, center, closed):
...:        start = np.empty(num_values, dtype=np.int64)
...:        end = np.empty(num_values, dtype=np.int64)
...:        for i in range(num_values):
...:            if self.use_expanding[i]:
...:                start[i] = 0
...:                end[i] = i + 1
...:            else:
...:                start[i] = i
...:                end[i] = i + self.window_size
...:        return start, end
...:

In [3]: indexer = CustomIndexer(window_size=1, use_expanding=use_expanding)

In [4]: df.rolling(indexer).sum()
Out[4]:
    values
0     0.0
1     1.0
2     3.0
3     3.0
4    10.0

You can view other examples of BaseIndexer subclasses here

New in version 1.1.

One subclass of note within those examples is the VariableOffsetWindowIndexer that allows rolling operations over a non-fixed offset like a BusinessDay.

In [32]: from pandas.api.indexers import VariableOffsetWindowIndexer

In [33]: df = pd.DataFrame(range(10), index=pd.date_range("2020", periods=10))

In [34]: offset = pd.offsets.BDay(1)

In [35]: indexer = VariableOffsetWindowIndexer(index=df.index, offset=offset)

In [36]: df
Out[36]: 
            0
2020-01-01  0
2020-01-02  1
2020-01-03  2
2020-01-04  3
2020-01-05  4
2020-01-06  5
2020-01-07  6
2020-01-08  7
2020-01-09  8
2020-01-10  9

In [37]: df.rolling(indexer).sum()
Out[37]: 
               0
2020-01-01   0.0
2020-01-02   1.0
2020-01-03   2.0
2020-01-04   3.0
2020-01-05   7.0
2020-01-06  12.0
2020-01-07   6.0
2020-01-08   7.0
2020-01-09   8.0
2020-01-10   9.0

For some problems knowledge of the future is available for analysis. For example, this occurs when each data point is a full time series read from an experiment, and the task is to extract underlying conditions. In these cases it can be useful to perform forward-looking rolling window computations. FixedForwardWindowIndexer class is available for this purpose. This BaseIndexer subclass implements a closed fixed-width forward-looking rolling window, and we can use it as follows:

Rolling apply

The apply() function takes an extra func argument and performs generic rolling computations. The func argument should be a single function that produces a single value from an ndarray input. raw specifies whether the windows are cast as Series objects (raw=False) or ndarray objects (raw=True).

In [38]: def mad(x):
   ....:     return np.fabs(x - x.mean()).mean()
   ....: 

In [39]: s = pd.Series(range(10))

In [40]: s.rolling(window=4).apply(mad, raw=True)
Out[40]: 
0    NaN
1    NaN
2    NaN
3    1.0
4    1.0
5    1.0
6    1.0
7    1.0
8    1.0
9    1.0
dtype: float64

Numba engine

New in version 1.0.

Additionally, apply() can leverage Numba if installed as an optional dependency. The apply aggregation can be executed using Numba by specifying engine='numba' and engine_kwargs arguments (raw must also be set to True). Numba will be applied in potentially two routines:

1. If func is a standard Python function, the engine will JIT the passed function. func can also be a JITed function in which case the engine will not JIT the function again.

2. The engine will JIT the for loop where the apply function is applied to each window.

The engine_kwargs argument is a dictionary of keyword arguments that will be passed into the numba.jit decorator. These keyword arguments will be applied to both the passed function (if a standard Python function) and the apply for loop over each window. Currently only nogil, nopython, and parallel are supported, and their default values are set to False, True and False respectively.

Note

In terms of performance, the first time a function is run using the Numba engine will be slow as Numba will have some function compilation overhead. However, the compiled functions are cached, and subsequent calls will be fast. In general, the Numba engine is performant with a larger amount of data points (e.g. 1+ million).

In [1]: data = pd.Series(range(1_000_000))

In [2]: roll = data.rolling(10)

In [3]: def f(x):
   ...:     return np.sum(x) + 5
# Run the first time, compilation time will affect performance
In [4]: %timeit -r 1 -n 1 roll.apply(f, engine='numba', raw=True)  # noqa: E225, E999
1.23 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
# Function is cached and performance will improve
In [5]: %timeit roll.apply(f, engine='numba', raw=True)
188 ms ± 1.93 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [6]: %timeit roll.apply(f, engine='cython', raw=True)
3.92 s ± 59 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Binary window functions

cov() and corr() can compute moving window statistics about two Series or any combination of DataFrame/Series or DataFrame/DataFrame. Here is the behavior in each case:

• two Series: compute the statistic for the pairing.

• DataFrame/Series: compute the statistics for each column of the DataFrame with the passed Series, thus returning a DataFrame.

• DataFrame/DataFrame: by default compute the statistic for matching column names, returning a DataFrame. If the keyword argument pairwise=True is passed then computes the statistic for each pair of columns, returning a MultiIndexed DataFrame whose index are the dates in question (see the next section).

For example:

In [41]: df = pd.DataFrame(
   ....:     np.random.randn(10, 4),
   ....:     index=pd.date_range("2020-01-01", periods=10),
   ....:     columns=["A", "B", "C", "D"],
   ....: )
   ....: 

In [42]: df = df.cumsum()

In [43]: df2 = df[:4]

In [44]: df2.rolling(window=2).corr(df2["B"])
Out[44]: 
              A    B    C    D
2020-01-01  NaN  NaN  NaN  NaN
2020-01-02 -1.0  1.0 -1.0  1.0
2020-01-03  1.0  1.0  1.0 -1.0
2020-01-04 -1.0  1.0  1.0 -1.0

Computing rolling pairwise covariances and correlations

In financial data analysis and other fields it’s common to compute covariance and correlation matrices for a collection of time series. Often one is also interested in moving-window covariance and correlation matrices. This can be done by passing the pairwise keyword argument, which in the case of DataFrame inputs will yield a MultiIndexed DataFrame whose index are the dates in question. In the case of a single DataFrame argument the pairwise argument can even be omitted:

Note

Missing values are ignored and each entry is computed using the pairwise complete observations. Please see the covariance section for caveats associated with this method of calculating covariance and correlation matrices.

In [45]: covs = (
   ....:     df[["B", "C", "D"]]
   ....:     .rolling(window=4)
   ....:     .cov(df[["A", "B", "C"]], pairwise=True)
   ....: )
   ....: 

In [46]: covs
Out[46]: 
                     B         C         D
2020-01-01 A       NaN       NaN       NaN
           B       NaN       NaN       NaN
           C       NaN       NaN       NaN
2020-01-02 A       NaN       NaN       NaN
           B       NaN       NaN       NaN
...                ...       ...       ...
2020-01-09 B  0.342006  0.230190  0.052849
           C  0.230190  1.575251  0.082901
2020-01-10 A -0.333945  0.006871 -0.655514
           B  0.649711  0.430860  0.469271
           C  0.430860  0.829721  0.055300

[30 rows x 3 columns]

Weighted window

The win_type argument in .rolling generates a weighted windows that are commonly used in filtering and spectral estimation. win_type must be string that corresponds to a scipy.signal window function. Scipy must be installed in order to use these windows, and supplementary arguments that the Scipy window methods take must be specified in the aggregation function.

In [47]: s = pd.Series(range(10))

In [48]: s.rolling(window=5).mean()
Out[48]: 
0    NaN
1    NaN
2    NaN
3    NaN
4    2.0
5    3.0
6    4.0
7    5.0
8    6.0
9    7.0
dtype: float64

In [49]: s.rolling(window=5, win_type="triang").mean()
Out[49]: 
0    NaN
1    NaN
2    NaN
3    NaN
4    2.0
5    3.0
6    4.0
7    5.0
8    6.0
9    7.0
dtype: float64

# Supplementary Scipy arguments passed in the aggregation function
In [50]: s.rolling(window=5, win_type="gaussian").mean(std=0.1)
Out[50]: 
0    NaN
1    NaN
2    NaN
3    NaN
4    2.0
5    3.0
6    4.0
7    5.0
8    6.0
9    7.0
dtype: float64

For all supported aggregation functions, see Weighted window functions.

Expanding window

An expanding window yields the value of an aggregation statistic with all the data available up to that point in time. Since these calculations are a special case of rolling statistics, they are implemented in pandas such that the following two calls are equivalent:

In [51]: df = pd.DataFrame(range(5))

In [52]: df.rolling(window=len(df), min_periods=1).mean()
Out[52]: 
     0
0  0.0
1  0.5
2  1.0
3  1.5
4  2.0

In [53]: df.expanding(min_periods=1).mean()
Out[53]: 
     0
0  0.0
1  0.5
2  1.0
3  1.5
4  2.0

For all supported aggregation functions, see Expanding window functions.

Exponentially Weighted window

An exponentially weighted window is similar to an expanding window but with each prior point being exponentially weighted down relative to the current point.

In general, a weighted moving average is calculated as

\[y_t = \frac{\sum_{i=0}^t w_i x_{t-i}}{\sum_{i=0}^t w_i},\]

where \(x_t\) is the input, \(y_t\) is the result and the \(w_i\) are the weights.

For all supported aggregation functions, see Exponentially-weighted window functions.

The EW functions support two variants of exponential weights. The default, adjust=True, uses the weights \(w_i = (1 - \alpha)^i\) which gives

\[y_t = \frac{x_t + (1 - \alpha)x_{t-1} + (1 - \alpha)^2 x_{t-2} + ... + (1 - \alpha)^t x_{0}}{1 + (1 - \alpha) + (1 - \alpha)^2 + ... + (1 - \alpha)^t}\]

When adjust=False is specified, moving averages are calculated as

\[\begin{split}y_0 &= x_0 \\ y_t &= (1 - \alpha) y_{t-1} + \alpha x_t,\end{split}\]

which is equivalent to using weights

\[\begin{split}w_i = \begin{cases} \alpha (1 - \alpha)^i & \text{if } i < t \\ (1 - \alpha)^i & \text{if } i = t. \end{cases}\end{split}\]

Note

These equations are sometimes written in terms of \(\alpha' = 1 - \alpha\), e.g.

\[y_t = \alpha' y_{t-1} + (1 - \alpha') x_t.\]

The difference between the above two variants arises because we are dealing with series which have finite history. Consider a series of infinite history, with adjust=True:

\[y_t = \frac{x_t + (1 - \alpha)x_{t-1} + (1 - \alpha)^2 x_{t-2} + ...} {1 + (1 - \alpha) + (1 - \alpha)^2 + ...}\]

Noting that the denominator is a geometric series with initial term equal to 1 and a ratio of \(1 - \alpha\) we have

\[\begin{split}y_t &= \frac{x_t + (1 - \alpha)x_{t-1} + (1 - \alpha)^2 x_{t-2} + ...} {\frac{1}{1 - (1 - \alpha)}}\\ &= [x_t + (1 - \alpha)x_{t-1} + (1 - \alpha)^2 x_{t-2} + ...] \alpha \\ &= \alpha x_t + [(1-\alpha)x_{t-1} + (1 - \alpha)^2 x_{t-2} + ...]\alpha \\ &= \alpha x_t + (1 - \alpha)[x_{t-1} + (1 - \alpha) x_{t-2} + ...]\alpha\\ &= \alpha x_t + (1 - \alpha) y_{t-1}\end{split}\]

which is the same expression as adjust=False above and therefore shows the equivalence of the two variants for infinite series. When adjust=False, we have \(y_0 = x_0\) and \(y_t = \alpha x_t + (1 - \alpha) y_{t-1}\). Therefore, there is an assumption that \(x_0\) is not an ordinary value but rather an exponentially weighted moment of the infinite series up to that point.

One must have \(0 < \alpha \leq 1\), and while it is possible to pass \(\alpha\) directly, it’s often easier to think about either the span, center of mass (com) or half-life of an EW moment:

\[\begin{split}\alpha = \begin{cases} \frac{2}{s + 1}, & \text{for span}\ s \geq 1\\ \frac{1}{1 + c}, & \text{for center of mass}\ c \geq 0\\ 1 - \exp^{\frac{\log 0.5}{h}}, & \text{for half-life}\ h > 0 \end{cases}\end{split}\]

One must specify precisely one of span, center of mass, half-life and alpha to the EW functions:

• Span corresponds to what is commonly called an “N-day EW moving average”.

• Center of mass has a more physical interpretation and can be thought of in terms of span: \(c = (s - 1) / 2\).

• Half-life is the period of time for the exponential weight to reduce to one half.

• Alpha specifies the smoothing factor directly.

New in version 1.1.0.

You can also specify halflife in terms of a timedelta convertible unit to specify the amount of time it takes for an observation to decay to half its value when also specifying a sequence of times.

In [54]: df = pd.DataFrame({"B": [0, 1, 2, np.nan, 4]})

In [55]: df
Out[55]: 
     B
0  0.0
1  1.0
2  2.0
3  NaN
4  4.0

In [56]: times = ["2020-01-01", "2020-01-03", "2020-01-10", "2020-01-15", "2020-01-17"]

In [57]: df.ewm(halflife="4 days", times=pd.DatetimeIndex(times)).mean()
Out[57]: 
          B
0  0.000000
1  0.585786
2  1.523889
3  1.523889
4  3.233686

The following formula is used to compute exponentially weighted mean with an input vector of times:

\[y_t = \frac{\sum_{i=0}^t 0.5^\frac{t_{t} - t_{i}}{\lambda} x_{t-i}}{0.5^\frac{t_{t} - t_{i}}{\lambda}},\]

ExponentialMovingWindow also has an ignore_na argument, which determines how intermediate null values affect the calculation of the weights. When ignore_na=False (the default), weights are calculated based on absolute positions, so that intermediate null values affect the result. When ignore_na=True, weights are calculated by ignoring intermediate null values. For example, assuming adjust=True, if ignore_na=False, the weighted average of 3, NaN, 5 would be calculated as

\[\frac{(1-\alpha)^2 \cdot 3 + 1 \cdot 5}{(1-\alpha)^2 + 1}.\]

Whereas if ignore_na=True, the weighted average would be calculated as

\[\frac{(1-\alpha) \cdot 3 + 1 \cdot 5}{(1-\alpha) + 1}.\]

The var(), std(), and cov() functions have a bias argument, specifying whether the result should contain biased or unbiased statistics. For example, if bias=True, ewmvar(x) is calculated as ewmvar(x) = ewma(x**2) - ewma(x)**2; whereas if bias=False (the default), the biased variance statistics are scaled by debiasing factors

\[\frac{\left(\sum_{i=0}^t w_i\right)^2}{\left(\sum_{i=0}^t w_i\right)^2 - \sum_{i=0}^t w_i^2}.\]

(For \(w_i = 1\), this reduces to the usual \(N / (N - 1)\) factor, with \(N = t + 1\).) See Weighted Sample Variance on Wikipedia for further details.

Group by: split-apply-combine Time series / date functionality