8

I've been told I need to normalise my MSE for my thesis involving neural networks.

Equations for NMSE seem a bit few and far-between. I have the following and want to corroborate it if possible:

enter image description here

Is the standard deviation term supposed to be calculated from the target values or the predicted values?

Also, what are the main advantages for using MSE over NMSE? Is it just that it makes error comparisons easier, because of the simpler scale?

Many thanks for any help!

Archie
  • 333
  • 2
  • 9
  • 1
    Standard deviation should be calculated from data, not the prediction. Advantages? I would say normalization allows you to compare distributions on a standardized scale, same for all. – duffymo Feb 19 '17 at 12:47
  • Table III in this paper defines it almost in the same way that you do but without the $100$ constant on the numerator https://ieeexplore.ieee.org/abstract/document/1257413/ – Rajiv Krishnakumar Sep 02 '22 at 08:51

1 Answers1

-2
    def nmser(x,y):
        z=0
        if len(x)==len(y):
            for k in range(len(x)):
                z = z + (((x[k]-y[k])**2)/x[k])    
                z = z/(len(x))
        return z
indhra
  • 7
  • 4