I noticed that you want startDate changed to the end of the respective 30-days period, starting from your start date.

Another detail is that you want the result grouped by INDEX_250 - how many previous transactions (in both respective periods) were with the same value of INDEX_250.

Note also that rolling calculations can be performed on a window containing rows from a number of future periods, whereas you want the number of transactions from previous 30 or 60 days and rolling does not allow negative number of periods.

This is why I took a different approach than "ordinary" rolling.

Start from auxiliary variables:

td30 = pd.Timedelta('30D')
dRng = pd.date_range(start='2013-09-17', end=df.startDate.max() + td30,
    freq='30D', closed='left')

Then define the following function computing both targets:

def targets(grp):
    grp['Prd'] = grp.startDate.apply(lambda x: dRng.asof(x) + td30)
    grp.set_index('Prd', inplace=True)
    trg30 = grp.groupby(level=0).INDEX_250.count()\
        .rename('nrTargets_gr_250_30').reindex(dRng, fill_value=0)
    trg60 = trg30.rolling(2).sum().rename('nrTargets_gr_250_60')\
        .fillna(0, downcast='infer')
    trg30 = trg30[trg30 > 0]
    trg60 = trg60[trg60 > 0]
    return trg30.to_frame().join(trg60, how='outer')\
        .fillna(0, downcast='infer').rename_axis('startDate')

Apply it and reset indices (just in this order, to have proper ordering of columns):

df2 = df[df.startDate >= '2013-09-17'].groupby('INDEX_250')\
    .apply(targets).reset_index(level=[0]).reset_index()

Notes:

• I took only rows with startDate on or after the start date you specified (2013-09-17).
• The type of both target columns is int. I think it is more natural, as these column contain the number of transactions, which by nature is just an integer.

And the last thing is to change the type of INDEX_250 to int:

df2.INDEX_250 = df2.INDEX_250.astype(int)

The result for INDEX_250 group is just as you specified, excluding result rows from 2016 and 2017, which were not included in your sample data.

Extended version - with average prices

To extend the result by average prices for each "final" date and both targets, two changes are required.

First, define another function to "reformat" target DataFrame:

def trgReformat(trg):
    trg = trg[trg.nrTargets_gr_250 > 0].copy()
    trg['avgPrice'] = trg.sm / trg.nrTargets_gr_250
    return trg.drop(columns='sm')

Second, define targets function as:

def targets(grp):
    grp['Prd'] = grp.startDate.apply(lambda x: dRng.asof(x) + td30)
    grp.set_index('Prd', inplace=True)
    trg30 = grp.groupby(level=0).agg(
        nrTargets_gr_250=('INDEX_250', 'count'), sm=('priceDeal', 'sum'))\
        .reindex(dRng, fill_value=0)
    trg60 = trg30.rolling(2).sum().fillna(0, downcast='infer')
    trg30 = trgReformat(trg30)
    trg60 = trgReformat(trg60)
    return trg30.join(trg60, how='outer', lsuffix='_30', rsuffix='_60')\
        .fillna(0, downcast='infer').rename_axis('startDate')

This function uses named aggregations, to compute:

• nrTargets_gr_250 - the number of rows,
• sm - sum of prices.

The reason is that computing of trg60 is performed using rolling (for 2 consecutive 30-day periods), so average alone would be not sufficient here.

Computation of the average price can be performed as late as in reformatting of each target.

Application of this function is just as before.