import pandas as pd

In [109]: s = pd.Series([1,2,3,4,6,3,4,0,2])

In [110]: s
Out[110]:
0    1
1    2
2    3
3    4
4    6
5    3
6    4
7    0
8    2
dtype: int64

In [111]: s.rolling(3, min_periods=1).sum()
Out[111]:
0     1.0
1     3.0
2     6.0
3     9.0
4    13.0
5    13.0
6    13.0
7     7.0
8     6.0
dtype: float64

In [112]: idx = s.rolling(3, min_periods=1).sum().idxmax()

In [113]: idx
Out[113]: 4

In [114]: s.loc[idx-2:idx]
Out[114]:
2    3
3    4
4    6
dtype: int64

In [18]: a = np.random.randint(10**4, size=10**6)

In [19]: a
Out[19]: array([9918, 4299, 7829, ..., 7513, 3367, 7140])

In [20]: pd.options.display.max_rows = 15

In [21]: s = pd.Series(a)

In [22]: %%timeit
    ...: idx = s.rolling(3, min_periods=1).sum().idxmax()
    ...: s.loc[idx-2:idx]
    ...:
115 ms ± 7.44 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [23]: s.shape
Out[23]: (1000000,)

$ python -mtimeit -s 'import numpy as np; a = np.random.randint(10**4, size=10**6); from max_rolling_sum import max_rolling_sum as f' 'f(a, 3)'
100 loops, best of 3: 2.7 msec per loop
$ python -mtimeit -s 'import numpy as np; a = np.random.randint(10**4, size=10**6); from max_rolling_maxu import max_rolling_sum as f' 'f(a, 3)'
10 loops, best of 3: 48.1 msec per loop

#cython: boundscheck=False

ctypedef long array_type

cpdef Py_ssize_t max_rolling_sum(array_type[:] arr, Py_ssize_t k) nogil:
    """arr[i:i+k].sum() is maximum."""
    cdef Py_ssize_t N = arr.shape[0]
    if N < 1:
        return -1  # error: no sum

    cdef Py_ssize_t i
    cdef array_type sum_ = 0
    for i in range(min(k, N)):  # find first sum arr[:k].sum()
        sum_ += arr[i]
    cdef Py_ssize_t max_start = 0
    cdef array_type max_sum = sum_
    for i in range(k, N):   # compute rolling sum arr[i-k+1:i+1].sum()
        sum_ -= arr[i - k]  # pop (left) from old sum
        sum_ += arr[i]      # append (right) to new sum
        if max_sum < sum_:
            max_sum = sum_
            max_start = i - k + 1
    return max_start

import pandas as pd

def max_rolling_sum(arr, k):
    """arr[i:i+k].sum() is maximum."""
    s = pd.Series(arr)
    idx = s.rolling(k, min_periods=1).sum().idxmax()
    return max(idx - 2, 0)