My suggestion would be not to group using pandas; but instead do it yourself. My guess is that the aggregation is implemented like a reduce, so theano is getting the equivalent of
tt.sum(a[0], tt.sum(a[1], tt.sum(a[2], ...
My suggestion is to convert the initial aggregated data frame (yydat – which initially contains only data so can use np.sum) into a dictionary keyed by your grouping variables, and to operate on the dictionary instead of using pandas functions.