Using Cumsum In Pandas On Group()

From a Pandas newbie: I have data that looks essentially like this - data1=pd.DataFrame({'Dir':['E','E','W','W','E','W','W','E'], 'Bool':['Y','N','Y','N','Y','N','Y','N'], 'Data':

Solution 1:

Try this:

data2 = data1.reset_index()
data3 = data2.set_index(["Bool", "Dir", "index"])   # index is the new column created by reset_indexrunning_sum = data3.groupby(level=[0,1,2]).sum().groupby(level=[0,1]).cumsum()

The reason you cannot simply use cumsum on data3 has to do with how your data is structured. Grouping by Bool and Dir and applying an aggregation function (sum, mean, etc) would produce a DataFrame of a smaller size than you started with, as whatever function you used would aggregate values based on your group keys. However cumsum is not an aggreagation function. It wil return a DataFrame that is the same size as the one it's called with. So unless your input DataFrame is in a format where the output can be the same size after calling cumsum, it will throw an error. That's why I called sum first, which returns a DataFrame in the correct input format.

Sorry if I haven't explained this well enough. Maybe someone else could help me out?

Solution 2:

As the other answer points out, you're trying to collapse identical dates into single rows, whereas the cumsum function will return a series of the same length as the original DataFrame. Stated differently, you actually want to group by [Bool, Dir, Date], calculate a sum in each group, THEN return a cumsum on rows grouped by [Bool, Dir]. The other answer is a perfectly valid solution to your specific question, here's a one-liner variation:

data1.groupby(['Bool', 'Dir', 'Date']).sum().groupby(level=[0, 1]).cumsum()

This returns output exactly in the requested format.

For those looking for a simple cumsum on a Pandas group, you can use:

data1.groupby(['Bool', 'Dir']).apply(lambda x: x['Data'].cumsum())

The cumulative sum is calculated internal to each group. Here's what the output looks like:

BoolDirNE2000-12-30     52000-12-30    16W2001-01-02     72001-01-03    16YE2000-12-30     42001-01-03    12W2000-12-30     62000-12-30    16Name:Data,dtype:int64

Note the repeated dates, but this is doing a strict cumulative sum internal to the rows of each group identified by the Bool and Dir columns.