polars.DataFrame.rolling#

DataFrame.rolling(
index_column: IntoExpr,
*,
period: str | timedelta,
offset: str | timedelta | None = None,
closed: ClosedInterval = 'right',
group_by: IntoExpr | Iterable[IntoExpr] | None = None,
) → RollingGroupBy[source]#

Create rolling groups based on a temporal or integer column.

Different from a group_by_dynamic the windows are now determined by the individual values and are not of constant intervals. For constant intervals use DataFrame.group_by_dynamic().

If you have a time series <t_0, t_1, ..., t_n>, then by default the windows created will be

  • (t_0 - period, t_0]

  • (t_1 - period, t_1]

  • …

  • (t_n - period, t_n]

whereas if you pass a non-default offset, then the windows will be

  • (t_0 + offset, t_0 + offset + period]

  • (t_1 + offset, t_1 + offset + period]

  • …

  • (t_n + offset, t_n + offset + period]

The period and offset arguments are created either from a timedelta, or by using the following string language:

  • 1ns (1 nanosecond)

  • 1us (1 microsecond)

  • 1ms (1 millisecond)

  • 1s (1 second)

  • 1m (1 minute)

  • 1h (1 hour)

  • 1d (1 calendar day)

  • 1w (1 calendar week)

  • 1mo (1 calendar month)

  • 1q (1 calendar quarter)

  • 1y (1 calendar year)

  • 1i (1 index count)

Or combine them: “3d12h4m25s” # 3 days, 12 hours, 4 minutes, and 25 seconds

By “calendar day”, we mean the corresponding time on the next day (which may not be 24 hours, due to daylight savings). Similarly for “calendar week”, “calendar month”, “calendar quarter”, and “calendar year”.

Parameters:
index_column

Column used to group based on the time window. Often of type Date/Datetime. This column must be sorted in ascending order (or, if group_by is specified, then it must be sorted in ascending order within each group).

In case of a rolling operation on indices, dtype needs to be one of {UInt32, UInt64, Int32, Int64}. Note that the first three get temporarily cast to Int64, so if performance matters use an Int64 column.

period

Length of the window - must be non-negative.

offset

Offset of the window. Default is -period.

closed{‘right’, ‘left’, ‘both’, ‘none’}

Define which sides of the temporal interval are closed (inclusive).

group_by

Also group by this column/these columns

Returns:
RollingGroupBy

Object you can call .agg on to aggregate by groups, the result of which will be sorted by index_column (but note that if group_by columns are passed, it will only be sorted within each group).

See also

group_by_dynamic

Examples

>>> dates = [
...     "2020-01-01 13:45:48",
...     "2020-01-01 16:42:13",
...     "2020-01-01 16:45:09",
...     "2020-01-02 18:12:48",
...     "2020-01-03 19:45:32",
...     "2020-01-08 23:16:43",
... ]
>>> df = pl.DataFrame({"dt": dates, "a": [3, 7, 5, 9, 2, 1]}).with_columns(
...     pl.col("dt").str.strptime(pl.Datetime).set_sorted()
... )
>>> out = df.rolling(index_column="dt", period="2d").agg(
...     [
...         pl.sum("a").alias("sum_a"),
...         pl.min("a").alias("min_a"),
...         pl.max("a").alias("max_a"),
...     ]
... )
>>> assert out["sum_a"].to_list() == [3, 10, 15, 24, 11, 1]
>>> assert out["max_a"].to_list() == [3, 7, 7, 9, 9, 1]
>>> assert out["min_a"].to_list() == [3, 3, 3, 3, 2, 1]
>>> out
shape: (6, 4)
┌─────────────────────┬───────┬───────┬───────┐
│ dt                  ┆ sum_a ┆ min_a ┆ max_a │
│ ---                 ┆ ---   ┆ ---   ┆ ---   │
│ datetime[μs]        ┆ i64   ┆ i64   ┆ i64   │
╞═════════════════════╪═══════╪═══════╪═══════╡
│ 2020-01-01 13:45:48 ┆ 3     ┆ 3     ┆ 3     │
│ 2020-01-01 16:42:13 ┆ 10    ┆ 3     ┆ 7     │
│ 2020-01-01 16:45:09 ┆ 15    ┆ 3     ┆ 7     │
│ 2020-01-02 18:12:48 ┆ 24    ┆ 3     ┆ 9     │
│ 2020-01-03 19:45:32 ┆ 11    ┆ 2     ┆ 9     │
│ 2020-01-08 23:16:43 ┆ 1     ┆ 1     ┆ 1     │
└─────────────────────┴───────┴───────┴───────┘

If you use an index count in period or offset, then it’s based on the values in index_column:

>>> df = pl.DataFrame({"int": [0, 4, 5, 6, 8], "value": [1, 4, 2, 4, 1]})
>>> df.rolling("int", period="3i").agg(pl.col("int").alias("aggregated"))
shape: (5, 2)
┌─────┬────────────┐
│ int ┆ aggregated │
│ --- ┆ ---        │
│ i64 ┆ list[i64]  │
╞═════╪════════════╡
│ 0   ┆ [0]        │
│ 4   ┆ [4]        │
│ 5   ┆ [4, 5]     │
│ 6   ┆ [4, 5, 6]  │
│ 8   ┆ [6, 8]     │
└─────┴────────────┘

If you want the index count to be based on row number, then you may want to combine rolling with with_row_index().

On this page