The simplest model for a neural net reference is just assuming that the output is a constant. The minimum MSE occurs when that constant is the target mean and MSE is the average target variance. Using that as a reference for more complex models yields
OUTref = mean(target')'
MSEref = mean(var(target',0))
the resulting normalized MSE for the general case is
NMSE = MSE/MSEref
The typical range is
0 <= NMSE <= 1
My typical goal for the hundreds of examples that I have posted is
NMSEgoal <= 0.01
or, for those familiar with the Rsquared parameter
0 <= Rsquared = 1-NMSE <= 0.99
Hope this helps
Greg
