Target and output in neural networks In ANN the output squeezed using sigmoid function so the result is always between 1 and  -1.
How am I supposed to calculate the error when the target value  might be  a big number?
For example I'm using ANN to predict a time series, so I get output like 0.98 but the target is 34.5.
 A: You can either:


*

*have no activation function on your output layer (the "identity" activation function).

*scale your targets so that their range is spanned by the activation function of the output layer. For a typical sigmoid (logistic activation function), it would be between 0 and 1 (not -1 and 1 as you wrote).


Usually the second option is preferred because it is easier to initialize the weights.
A: The sigmoid activation is good for outputs bounded between 0 and 1 such as probabilities for classification. 
In case you don't need specific bounds you can use other non-linearities like ReLU(x) = max(0,x). In case you do know some bounds for the output you can use the sigmoid and then do a linear transformation.
Warning: I saw that you are trying to train a Neural Network for a time series. Note that if you're starting with Neural Networks you may want to start with a simple FeedForward network applied to some simple problem (sequences are usually handled with Recurrent Neural Nets which are harder to train).
