Dear S-plus users,
I have designed a simple backpropagation neural network to perform
time series prediction, that is, inputs x(t), x(t-1),..., x(t-N+1) must
produce x(t+1) in the output. Training converges, but when I pass the test
set through the network, the output is delayed by one sample (that is,
the output follows x(t) and not (x(t+1)). Also surprising is the fact
that, when I train a network to generate x(t+1) and x(t+2), both outputs
follow x(t). This behavior is valid only for the test set; the training
and validation sets are correctly mapped.
Has anyone seen this before? Any help would be truly welcome!
Regards,
Augusto.
--
Augusto Cesar Heluy Dantas
D.Sc. Student at COPPE/UFRJ
Rio de Janeiro, Brazil
|