I'm new to forecasting using neural networks. I have decided to use feedforward backpropagation algorithm. What are the input values if I have past data and what is the technique to select input values?
$\begingroup$
$\endgroup$
1
-
$\begingroup$ Can you please add some details? Now it is like asking in which shop you should buy ingredients provided that you decided to use copper pan. $\endgroup$– user88Commented Jul 25, 2012 at 22:50
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
From what I understand, you are looking for time series prediction using a feed forward neural network. If that is the case, what you are interested in is to make the neural network learn the relation between the past values and the present value. Let me put that more formally, if ${y_1, y_2,...,y_T}$ is your input time series, at any time $t$ you would like to predict the next observation, i.e., $y_{t+1}$. In other words, you want to learn the function $f$ in $y_{t+1} = f(y_{t},y_{t-1},...y_{t-s})$.
Now, for that the input to the neural network has to be a delayed versions of the past 's' samples and the desired has to be ${y_{t+1}}$. Once you train the network, at each point you can 'forecast' the next observation.
$\begingroup$
$\endgroup$
I'm sure I'm not qualified to answer your question fully, but the way I have seen it done before is using something along the lines of a z-score for the factor that you are inputting.
So for example if you were inputting something like amount of bodyfat on a person, that has a mean of 40 pounds with standard deviation 10, for someone who has 50 pounds of fat, you would input the value 1 into the neural network.
answered Jul 25, 2012 at 20:10