Search code examples
pythondataframepython-polars

Performing integer-based rolling window group_by using Python Polars


I have a outer/inner loop-based function I'm trying to vectorise using Python Polars DataFrames. The function is a type of moving average and will be used to filter time-series financial data. Here's the function:

def ma_j(df_src: pl.DataFrame, depth: float):

    jrc04 = 0.0
    jrc05 = 0.0
    jrc06 = 0.0
    jrc08 = 0.0

    series = df_src['close']

    for x in range(0, len(series)):
        if x >= x - depth*2:
            for k in np.arange(start=math.ceil(depth), stop=0, step=-1):
                jrc04 = jrc04 + abs(series[x-k] - series[x-(k+1)])
                jrc05 = jrc05 + (depth + k) * abs(series[x-k] - series[x-(k+1)])
                jrc06 = jrc06 + series[x-(k+1)]
        else:
            jrc03 = abs(series - (series[1]))
            jrc13 = abs(series[x-depth] - series[x - (depth+1)])
            jrc04 = jrc04 - jrc13 + jrc03
            jrc05 = jrc05 - jrc04 + jrc03 * depth
            jrc06 = jrc06 - series[x - (depth+1)] + series[x-1]
        jrc08 = abs(depth * series[x] - jrc06)

    if jrc05 == 0.0:
        ma = 0.0
    else:
        ma = jrc08/jrc05

    return ma

The tricky bit for me are multiple the inner loop look-backs (for k in...). I've looked through multiple examples that use group_by_dynamic on the timeseries data. For example, here. I've also seen an example for rolling, but this still seems to use a period.

However, I'd like to strip away the timeseries and just use source Series. Does this mean I need to group on an integer range?

Using this data example:

import polars as pl
import numpy as np

i, t, v = np.arange(0, 50, 1), np.arange(0, 100, 2), np.random.randint(1,101,50)
df = pl.DataFrame({"i": i, "t": t, "rand": v})
df = df.with_columns((pl.datetime(2022,10,30) + pl.duration(seconds=df["t"])).alias("datetime")).drop("t")
cols = ["i", "datetime", "rand"]
df = df.select(cols)

DataFrame looks like this:

shape: (50, 3)
┌─────┬─────────────────────┬──────┐
│ i   ┆ datetime            ┆ rand │
│ --- ┆ ---                 ┆ ---  │
│ i64 ┆ datetime[μs]        ┆ i64  │
╞═════╪═════════════════════╪══════╡
│ 0   ┆ 2022-10-30 00:00:00 ┆ 87   │
│ 1   ┆ 2022-10-30 00:00:02 ┆ 66   │
│ 2   ┆ 2022-10-30 00:00:04 ┆ 30   │
│ 3   ┆ 2022-10-30 00:00:06 ┆ 87   │
│ 4   ┆ 2022-10-30 00:00:08 ┆ 74   │
│ …   ┆ …                   ┆ …    │
│ 45  ┆ 2022-10-30 00:01:30 ┆ 91   │
│ 46  ┆ 2022-10-30 00:01:32 ┆ 52   │
│ 47  ┆ 2022-10-30 00:01:34 ┆ 68   │
│ 48  ┆ 2022-10-30 00:01:36 ┆ 26   │
│ 49  ┆ 2022-10-30 00:01:38 ┆ 99   │
└─────┴─────────────────────┴──────┘

...I can do a grouping by datetime like this":

df.group_by_dynamic("datetime", every="10s").agg(
    pl.col("rand").mean().alias('rolling mean')
)

which gives this: enter image description here

But there's 3 issues with this:

  • I don't want to group of datetime...I want to group on every row (maybe i?) in bins of [x] size.
  • I need values against every row
  • I would like to define the aggregation function, as per the various cases in the function above

Any tips on how I could attack this using Polars? Thanks.

---------- Edit 1

Following @ritchie46 's awesome advice (thanks mate!), here's the groupby:

result_grp = (
    df
    .rolling(index_column="i", period="10i")
    .agg(
        pl.len().alias("rolling_slots"),
        pl.col("rand").mean().alias("roll_mean")
    )
)

df2 = df.select(
    pl.all(),
    result_grp.get_column("rolling_slots"),
    result_grp.get_column("roll_mean"),
)

This now gives:

shape: (50, 5)
┌─────┬─────────────────────┬──────┬───────────────┬───────────┐
│ i   ┆ datetime            ┆ rand ┆ rolling_slots ┆ roll_mean │
│ --- ┆ ---                 ┆ ---  ┆ ---           ┆ ---       │
│ i64 ┆ datetime[μs]        ┆ i64  ┆ u32           ┆ f64       │
╞═════╪═════════════════════╪══════╪═══════════════╪═══════════╡
│ 0   ┆ 2022-10-30 00:00:00 ┆ 23   ┆ 1             ┆ 23.0      │
│ 1   ┆ 2022-10-30 00:00:02 ┆ 72   ┆ 2             ┆ 47.5      │
│ 2   ┆ 2022-10-30 00:00:04 ┆ 46   ┆ 3             ┆ 47.0      │
│ 3   ┆ 2022-10-30 00:00:06 ┆ 37   ┆ 4             ┆ 44.5      │
│ 4   ┆ 2022-10-30 00:00:08 ┆ 12   ┆ 5             ┆ 38.0      │
│ …   ┆ …                   ┆ …    ┆ …             ┆ …         │
│ 45  ┆ 2022-10-30 00:01:30 ┆ 95   ┆ 10            ┆ 53.7      │
│ 46  ┆ 2022-10-30 00:01:32 ┆ 100  ┆ 10            ┆ 62.7      │
│ 47  ┆ 2022-10-30 00:01:34 ┆ 6    ┆ 10            ┆ 62.2      │
│ 48  ┆ 2022-10-30 00:01:36 ┆ 27   ┆ 10            ┆ 56.5      │
│ 49  ┆ 2022-10-30 00:01:38 ┆ 33   ┆ 10            ┆ 54.5      │
└─────┴─────────────────────┴──────┴───────────────┴───────────┘

This is great; now instead of mean(), how do I apply a custom function on the grouped values, such as:

f_jparams(depth_array, jrc04, jrc05, jrc06, jrc08):
    
    _depth = len(depth_array)
    
    if len(depth_array) > 3:
        for x in np.arange(start=1, stop=len(depth_array), step=1):
            jrc04 = jrc04 + abs(depth_array[x] - depth_array[x-1])
            jrc05 = jrc05 + (_depth+x) * abs(depth_array[x] - depth_array[x-1])
            jrc06 = jrc06 + depth_array[x-1]
    else:
        jrc03 = abs(depth_array[_depth-1] - depth_array[_depth-2])
        jrc13 = abs(depth_array[0] - depth_array[1])
        jrc04 = jrc04 - jrc13 + jrc03
        jrc05 = jrc05 - jrc04 + jrc03*_depth
        jrc06 = jrc06 - depth_array[1] + depth_array[_depth-2]
        
    jrc08 = abs(_depth * depth_array[0] - jrc06)
    
    if jrc05 == 0.0:
        ma = 0.0
    else:
        ma = jrc08/jrc05
        
    return ma, jrc04, jrc05, jrc06, jrc08

Thanks!

---- Edit 2:

Thanks to this post, I can collect up the items in the rand rolling group into a list for each row:

depth = 10

result_grp = (
    df
    .rolling(
        index_column="i", 
        period=str(depth) + "i",
        # offset="0i",
        # closed="left"
    )
    .agg(
        pl.len().alias("rolling_slots"),
        pl.col("rand").mean().alias("roll_mean"),
        pl.col("rand").name.suffix('_val_list'),
    )
)

df2 = df.select(
    pl.all(),
    result_grp.get_column("rolling_slots"),
    result_grp.get_column("roll_mean"),
    result_grp.get_column("rand_val_list"),
)

Also from this post, I saw a way to make the rolling window period a variable; nice!

Is there a way to use get_columns and exclude together so I don't have to list every col I want?

The dataframe now looks like:

shape: (50, 6)
┌─────┬─────────────────────┬──────┬───────────────┬───────────┬─────────────────┐
│ i   ┆ datetime            ┆ rand ┆ rolling_slots ┆ roll_mean ┆ rand_val_list   │
│ --- ┆ ---                 ┆ ---  ┆ ---           ┆ ---       ┆ ---             │
│ i64 ┆ datetime[μs]        ┆ i64  ┆ u32           ┆ f64       ┆ list[i64]       │
╞═════╪═════════════════════╪══════╪═══════════════╪═══════════╪═════════════════╡
│ 0   ┆ 2022-10-30 00:00:00 ┆ 23   ┆ 1             ┆ 23.0      ┆ [23]            │
│ 1   ┆ 2022-10-30 00:00:02 ┆ 72   ┆ 2             ┆ 47.5      ┆ [23, 72]        │
│ 2   ┆ 2022-10-30 00:00:04 ┆ 46   ┆ 3             ┆ 47.0      ┆ [23, 72, 46]    │
│ 3   ┆ 2022-10-30 00:00:06 ┆ 37   ┆ 4             ┆ 44.5      ┆ [23, 72, … 37]  │
│ 4   ┆ 2022-10-30 00:00:08 ┆ 12   ┆ 5             ┆ 38.0      ┆ [23, 72, … 12]  │
│ …   ┆ …                   ┆ …    ┆ …             ┆ …         ┆ …               │
│ 45  ┆ 2022-10-30 00:01:30 ┆ 95   ┆ 10            ┆ 53.7      ┆ [10, 11, … 95]  │
│ 46  ┆ 2022-10-30 00:01:32 ┆ 100  ┆ 10            ┆ 62.7      ┆ [11, 84, … 100] │
│ 47  ┆ 2022-10-30 00:01:34 ┆ 6    ┆ 10            ┆ 62.2      ┆ [84, 53, … 6]   │
│ 48  ┆ 2022-10-30 00:01:36 ┆ 27   ┆ 10            ┆ 56.5      ┆ [53, 46, … 27]  │
│ 49  ┆ 2022-10-30 00:01:38 ┆ 33   ┆ 10            ┆ 54.5      ┆ [46, 56, … 33]  │
└─────┴─────────────────────┴──────┴───────────────┴───────────┴─────────────────┘

Should I just now resort back to looping through the rand_val_list column and send each grouped values list to my function? Or is there a better polars way?

Thanks again!


Solution

  • Are you searching for periods="10i"?

    Polars rolling accepts a period argument with the following query language:

            - 1ns   (1 nanosecond)
            - 1us   (1 microsecond)
            - 1ms   (1 millisecond)
            - 1s    (1 second)
            - 1m    (1 minute)
            - 1h    (1 hour)
            - 1d    (1 day)
            - 1w    (1 week)
            - 1mo   (1 calendar month)
            - 1y    (1 calendar year)
            - 1i    (1 index count)
    

    Where i is simply the number of indices/rows.

    So on your data a rolling group_by where we count the number of slots would give:

    (df.rolling(index_column="i", period="10i")
       .agg(
           pl.len().alias("rolling_slots")
       )
    )
    
    shape: (50, 2)
    ┌─────┬───────────────┐
    │ i   ┆ rolling_slots │
    │ --- ┆ ---           │
    │ i64 ┆ u32           │
    ╞═════╪═══════════════╡
    │ 0   ┆ 1             │
    │ 1   ┆ 2             │
    │ 2   ┆ 3             │
    │ 3   ┆ 4             │
    │ 4   ┆ 5             │
    │ …   ┆ …             │
    │ 45  ┆ 10            │
    │ 46  ┆ 10            │
    │ 47  ┆ 10            │
    │ 48  ┆ 10            │
    │ 49  ┆ 10            │
    └─────┴───────────────┘