Pandas rolling sum with groupby and conditions

I have a dataframe with a timeseries of sales of different items with customer analytics. For each item and a given day I want to compute:

• a share of my best customer in last 2 days total sales
• a share of my top customers (from a list) in last 2 days total sales

I’ve tried solutions provided here:

• for rolling sum: Pandas sum over a date range for each category separately
• for conditioned groupby: Pandas groupby with identification of an element with max value in another column

An example dataframe is can be generated by:

import pandas as pd
from datetime import timedelta

df_1 = pd.DataFrame()
df_2 = pd.DataFrame()
df_3 = pd.DataFrame()

# Create datetimes and data
df_1['item'] = [1, 1, 1, 2, 2, 2, 2]
df_1['date'] = pd.date_range('1/1/2018', periods=7, freq='D')
df_1['customer'] = ['a', 'b', 'c', 'a', 'b', 'b', 'c']
df_1['sales'] = [2, 4, 1, 5, 7, 2, 3]

df_2['item'] = [1, 1, 1, 2, 2, 2, 2]
df_2['date'] = pd.date_range('1/1/2018', periods=7, freq='D')
df_2['customer'] = ['b', 'b', 'c', 'a', 'a', 'c', 'a']
df_2['sales'] = [2, 3, 4, 2, 3, 5, 6]

df_3['item'] = [1, 1, 1, 2, 2, 2, 2]
df_3['date'] = pd.date_range('1/1/2018', periods=7, freq='D')
df_3['customer'] = ['b', 'c', 'a', 'c', 'b', 'a', 'b']
df_3['sales'] = [6, 5, 2, 3, 4, 5, 6]

df = pd.concat([df_1, df_2, df_3])
df = df.sort_values(['item', 'date'])
df.reset_index(drop=True)

and looks like this:

item	date	customer	sales
1	2018-01-01	a	2
1	2018-01-01	b	2
1	2018-01-01	b	6
1	2018-01-02	b	4
1	2018-01-02	b	3
1	2018-01-02	c	5
1	2018-01-03	c	1
1	2018-01-03	c	4
1	2018-01-03	a	2
2	2018-01-04	a	5
2	2018-01-04	a	2
2	2018-01-04	c	3
2	2018-01-05	b	7
2	2018-01-05	a	3
2	2018-01-05	b	4
2	2018-01-06	b	2
2	2018-01-06	c	5
2	2018-01-06	a	5
2	2018-01-07	c	3
2	2018-01-07	a	6
2	2018-01-07	b	6

I expect the following results:

item	date	sales_at_day	sales_last_2_days	a_share	top_share
1	2018-01-01	10	NaN	NaN	NaN
1	2018-01-02	12	10	0.20	0.20
1	2018-01-03	7	22	0.09	0.09
2	2018-01-04	10	NaN	NaN	NaN
2	2018-01-05	14	10	0.70	1.00
2	2018-01-06	12	24	0.29	0.42
2	2018-01-07	15	26	0.31	0.50

where,

a_share is the share of sales of customer ‘a’ in total sales in last 2 days (not including present day) top_share is the share of sales of customers in a

top_cust = ['a', 'c'] 

list in total sales in last 2 days (not including present day)

Any ideas? Many Thanks in advance :)

Andy

Answer

Use:

#custom rolling with shift first day
f = lambda x: x.rolling(2, min_periods=1).sum().shift()

#aggregate sum
df1 = df.groupby(['item','date'], as_index=False)['sales'].sum()
#apply custom rolling per groups
df1['sales_last_2_days'] = df1.groupby('item')['sales'].apply(f).reset_index(drop=True, level=0)

#filter customer a and aggregate sum
a = df[df['customer'].eq('a')].groupby(['item','date'])['sales'].sum().rename('a_share')
#add new column to original
df1 = df1.join(a, on=['item','date'])
#applt custom rolling per groups and divide
df1['a_share'] = df1.groupby('item')['a_share'].apply(f).reset_index(drop=True, level=0) / df1['sales_last_2_days']

#verys similar like before, only test membership by isin
top_cust = ['a', 'c'] 
a = df[df['customer'].isin(top_cust)].groupby(['item','date'])['sales'].sum().rename('top_share')
df1 = df1.join(a, on=['item','date'])
df1['top_share'] = df1.groupby('item')['top_share'].apply(f).reset_index(drop=True, level=0) / df1['sales_last_2_days']
print (df1)
   item       date  sales  sales_last_2_days   a_share  top_share
0     1 2018-01-01     10                NaN       NaN        NaN
1     1 2018-01-02     12               10.0  0.200000   0.200000
2     1 2018-01-03      7               22.0  0.090909   0.318182
3     2 2018-01-04     10                NaN       NaN        NaN
4     2 2018-01-05     14               10.0  0.700000   1.000000
5     2 2018-01-06     12               24.0  0.416667   0.541667
6     2 2018-01-07     15               26.0  0.307692   0.500000

If want use rolling with days, it is more complicated:

df1 = df.groupby(['item','date'], as_index=False)['sales'].sum()
sales1 = (df1.set_index('date')
             .groupby('item')['sales']
             .rolling('2D', min_periods=1)
             .sum()
             .groupby('item')
             .shift()
             .rename('sales_last_2_days')
         )
df1 = df1.join(sales1, on=['item','date'])

---

df2 = df[df['customer'].eq('a')].groupby(['item','date'], as_index=False)['sales'].sum()
a = (df2.set_index('date')
        .groupby('item')[['sales']]
        .apply(lambda x: x.asfreq('D'))
        .reset_index(level=0)
        .groupby('item')['sales']
        .rolling('2D', min_periods=1)
        .sum()
        .groupby('item')
        .shift()
        .rename('a_share')
         )
print (a)
df1 = df1.join(a, on=['item','date'])
df1['a_share'] /= df1['sales_last_2_days']

---

top_cust = ['a', 'c'] 

df3 = df[df['customer'].isin(top_cust)].groupby(['item','date'], as_index=False)['sales'].sum()
b = (df3.set_index('date')
        .groupby('item')[['sales']]
        .apply(lambda x: x.asfreq('D'))
        .reset_index(level=0)
        .groupby('item')['sales']
        .rolling('2D', min_periods=1)
        .sum()
        .groupby('item')
        .shift()
        .rename('top_share')
         )
df1 = df1.join(b, on=['item','date'])
df1['top_share'] /= df1['sales_last_2_days']

---

print (df1)
   item       date  sales  sales_last_2_days   a_share  top_share
0     1 2018-01-01     10                NaN       NaN        NaN
1     1 2018-01-02     12               10.0  0.200000   0.200000
2     1 2018-01-03      7               22.0  0.090909   0.318182
3     2 2018-01-04     10                NaN       NaN        NaN
4     2 2018-01-05     14               10.0  0.700000   1.000000
5     2 2018-01-06     12               24.0  0.416667   0.541667
6     2 2018-01-07     15               26.0  0.307692   0.500000