There is no simple way to do that, because the argument that is passed to the rolling-applied function is a plain numpy array, not a pandas Series, so it doesn’t know about the index. Moreover, the rolling functions must return a float result, so they can’t directly return the index values if they’re not floats.
Here is one approach:
>>> s.index[s.rolling(3).apply(np.argmax)[2:].astype(int)+np.arange(len(s)-2)]
Index([u'c', u'c', u'e', u'e', u'e', u'f', u'i', u'i'], dtype="object")
The idea is to take the argmax values and align them with the series by adding a value indicating how far along in the series we are. (That is, for the first argmax value we add zero, because it is giving us the index into a subsequence starting at index 0 in the original series; for the second argmax value we add one, because it is giving us the index into a subsequence starting at index 1 in the original series; etc.)
This gives the correct results, but doesn’t include the two “None” values at the beginning; you’d have to add those back manually if you wanted them.
There is an open pandas issue to add rolling idxmax.