Does ANN really needs pre-process?

Many papers suggest to do certain preprocessing steps to improve ANN performance. The common way is the normalization of input and output data.

I currently use function fitting tool. I know this tool includes the mapminmax functions to normalize data to [-1,1]. I will not need to perforce normalization directly. Because the training results do not satisfy me. So I try to disable matlab default process function by

net.inputs{1}.processFcns ={};
net.outputs{1}.processFcns ={};

and do normalization by myself. In this way, all data are between [-1,1]. Before I do post-processing, the predicted outputs should fall between [-1,1], I think. However, sometimes I found the results are beyond [-1,1]. What is wrong?

[Walter Roberson]

It sounds as if you are asking what is wrong with your own normalization code, which you do not show us ?