Re: model selection & residuals

Dnia Fri, 28 Sep 2007 08:37:54 -0700
dave@xxxxxxxxxxx napisał(a):


You should be testing the statistical characteristics of the residuals
from the model and if they are independent and normal and identically
distributed THEN you should be reviewing the t values for each
estimated coefficient AND if that is all OK THEN and only then should
you use the model and it's parameters to obtain a set of forecasts.

I am afraid, Dave, but you know nothing about neural networks.

To the original poster:
- testing independence of residuals makes sense for
the situation you use NN for predicting time series
- I suggest repeating cross-validation and selecting models
basing on the average results
- assumption that residuals are normally distributed is
useful, mainly from the theoretical point of view. Most of
the cases are the linear models (eg. multiple regression,
time series models like arma).
- there is a lot of literature on training and selecting NN
models.

Best regards
--
[ Wit Jakuczun <W.Jakuczun [at] wlogsolutions.com> ]
[ WLOG Solutions http://www.wlogsolutions.com ]
.

Relevant Pages

  • Re: Time Series Regression Analysis
    ... Perhaps your colleague was referring to autocorrelation of residuals ... a common problem with time series models. ... Regression tool, and my book "Data Analysis Using Microsoft Excel: ...
    (microsoft.public.excel.misc)
  • autocorrelation as a predicor in multiple regression models
    ... I have got time series data (900 sec with a sampling rate of 0.1; ... The residuals are highly autocorrelated and so the ... assumptions of the regression model are violated and the estimates are ...
    (sci.stat.math)
  • Portmanteau test of residuals
    ... good lag value for the Portmanteau test (done on the residuals of a ... fitted model). ... documentation for time series of around n=100 or n=150, ...
    (sci.stat.math)