model selection & residuals

Hi,

I' m working with neural networks and I select my models based on a
ten-fold-crossvalidation and selection on the smallest RMSE of the
predicted and targeted (observed) values.
When I check the residuals in some cases (for some neural networks)
they are normally distributed and in other cases not. In some cases
this is not necessarely the network with the lowest RMSE.

Is it statistically justified to select the 'best' model based on two
selection criteria:
- low RMSE
- AND normally distributed residuals.

And thus leave out the models where the residuals are not normally
distributed. Or should the selection be based only on the RMSE,
independently of the normal distribution of the residuals.

I hope I've made myself and my question a little bit clear :)
Hope to hear from you,
Bea

.

Relevant Pages

  • Re: model selection & residuals
    ... ten-fold-crossvalidation and selection on the smallest RMSE of the ... When I check the residuals in some cases (for some neural networks) ... AND normally distributed residuals. ... independently of the normal distribution of the residuals. ...
    (sci.stat.math)
  • Re: model selection & residuals
    ... ten-fold-crossvalidation and selection on the smallest RMSE of the ... this is not necessarely the network with the lowest RMSE. ... AND normally distributed residuals. ... independently of the normal distribution of the residuals. ...
    (sci.stat.math)
  • Re: model selection & residuals
    ... ten-fold-crossvalidation and selection on the smallest RMSE of the ... When I check the residuals in some cases (for some neural networks) ... AND normally distributed residuals. ... independently of the normal distribution of the residuals. ...
    (sci.stat.math)