Very incorrect prediction

aoneil

I'm currently using a Neural Net with a series to then make a prediction for the next day. I pass the neural net five variables (though some are special), day (id), frequency/day (label), id_number (cluster), week, and day_of_week. Looking at the models after they've been built, I understand that it's taking week and day of week into account when making predictions (as it should). However, the frequency/day of every single example ranges from 0-600, yet when I send the model unlabeled information for one day in the future (day, week, day of week, id number, and date), the model will often make outrageous predictions such as a frequency of -20. Am I not setting this up correctly (too few features, etc), or am I just using the wrong type of model?

Thanks in advance for any input