You should describe your field more precisely. It's hard to suggest something, when you need to predict "array with variable length". For example, time-series prediction uses it's specific technics. So, pls give some info about your data.

Also, for regression you'd keep data normalized. If you use -100 in output, while other values would be very small like 0.1 it may lead to exploding gradient problem. Big values will affect on loss fn very much. So, if your goal is predicting vector try to use zero-padding.