Pandas date index of latest non null value in grouped date rolling

user_10011330 0
Viewed 274

I am trying to get the latest date at which a value was not null on a rolling time window, by group. It works pretty well without groups but it seems that the grouping shuffles everything.

Here is reproducible example:

import pandas as pd
from datetime import datetime as dt
import numpy as np

df = pd.DataFrame({})

df["date"] = [dt(2020, 10, i+1) for i in range(10)]
df["group"] = ["a" if int(i/3) == (i/3) else "b" for i in range(10)]
df["value"] = [i if int(i/2) == (i/2) else np.nan for i in range(10)]

dataframe

        date group  value
0 2020-10-01     a    0.0
1 2020-10-02     b    NaN
2 2020-10-03     b    2.0
3 2020-10-04     a    NaN
4 2020-10-05     b    4.0
5 2020-10-06     b    NaN
6 2020-10-07     a    6.0
7 2020-10-08     b    NaN
8 2020-10-09     b    8.0
9 2020-10-10     a    NaN

Targeted output:

        date group  value  output
0 2020-10-01     a    0.0  2020-10-01
1 2020-10-02     b    NaN  NaT
2 2020-10-03     b    2.0  2020-10-03
3 2020-10-04     a    NaN  2020-10-01
4 2020-10-05     b    4.0  2020-10-05
5 2020-10-06     b    NaN  2020-10-05
6 2020-10-07     a    6.0  2020-10-07
7 2020-10-08     b    NaN  2020-10-05
8 2020-10-09     b    8.0  2020-10-09
9 2020-10-10     a    NaN  2020-10-07

My attempt:

df = df.set_index("date").sort_index(ascending = True)

def latest_non_null_value_index(x):
        y = x[np.isnan(x) == False]
        print(y.index)
        if len(y) > 0:
            return y.index[-1]
        else:
            return np.nan

latest_index = df\
        .groupby(["group"])\
        .rolling("35D")\
        ["value"]\
        .apply(lambda x: latest_non_null_value_index(x).timestamp())\
        .reset_index()
  
def to_datetime_from_timestamp(x):
  if pd.isnull(x) == False:
      return dt.fromtimestamp(x)
  else:
      return pd.NaT
           
latest_index["value"] = latest_index["value"]\
    .apply(to_datetime_from_timestamp)

What I get:

  group       date               value
0     a 2020-10-01 2020-10-01 02:00:00
1     a 2020-10-04 2020-10-01 02:00:00
2     a 2020-10-07 2020-10-03 02:00:00
3     a 2020-10-10 2020-10-03 02:00:00
4     b 2020-10-02                 NaT
5     b 2020-10-03 2020-10-06 02:00:00
6     b 2020-10-05 2020-10-07 02:00:00
7     b 2020-10-06 2020-10-07 02:00:00
8     b 2020-10-08 2020-10-07 02:00:00
9     b 2020-10-09 2020-10-10 02:00:00

Any idea what I have missed here?

EDIT: also it seems that I don't have this problem when taking the latest value... It really has to do with index.

EDIT2: also if I could somehow apply a function to 2 columns, I could have the date as a second column and get a workaround

rolling-computation pandas group-by date python
1 Answers

You could use pd.fillna with the "ffill" to forward fill the missing values

import pandas as pd
from datetime import datetime as dt
import numpy as np

df = pd.DataFrame({})

df["date"] = [dt(2020, 10, i+1) for i in range(10)]
df["group"] = ["a" if int(i/3) == (i/3) else "b" for i in range(10)]
df["value"] = [i if int(i/2) == (i/2) else np.nan for i in range(10)]

df = df.sort_values("date")  # Just make sure that row are properly ordered

date = df["date"].copy()
date[df.value.isna()] = pd.NaT
latest_index = date.groupby(df.group).fillna(method="ffill")

This doesn't take care of your rolling time frame but you could remove the values that are outside of the time window like this:

latest_index[(df.date - latest_index).dt.days > 35] = pd.NaT

But that's not super tidy so you could try using the max aggregation against a rolling window like this:

df = df.set_index("date", drop=False)
df = df.sort_index()

date = pd.to_numeric(df["date"].copy())  # it wasn't letting me aggregate dates so we have to convert to float then back to dates
date[df.value.isna()] = None
latest_index = date.groupby(df.group).rolling("35D").max()
latest_index = pd.to_datetime(latest_index)
user_123989860
Related