Could you please help me understand this result? NARX net with removedelay -> unexpected output

Hello,

Experience: I have about 9 months experience with matlab R2015a and NN toolbox. Feeling fairly competent and certainly not new to programming.

Whats working: I have a trained NarxNet in which the output highly correlates to the target series. dataset is 3000+ samples.

Whats NOT working (and I'd love to know why): Once I remove one delay and go over the previously trained dataset in smaller chunks (~300). I grab the latest (one step ahead, e.g. 301) outputs, save those. Advance the smaller chunk window one sample and repeat. I would expect those one-step ahead predictions to be highly correlated to the original (3000+) data except shifted one step advanced. This is NOT the case. I attached a figure with an example, the blue is the Actual Target, Red is the NARX output with one delay removed, and Gold is the Actual Target manually advanced one step. Under my understanding of removing a delay, I would expect the Red to be correlated to the Gold.

Answer 1

[Greg Heath]

I have found the removedelay command USELESS and have recommended it's nonuse in at least one of my tutorials (NEWSGROUP and/or ANSWERS)

Hope this helps

AND

Hope you didn't waste as much time as I did trying to get it to work.

Greg

[Sarun777]

I have wasted a day trying to figure out removedelay. Thanks for the clear up. Now I just adjust the input dataset myself.