Using NN for forcasting

Hi all,

I would like to design Feed forward NN for stock price forecasting. I've 3 years stock price datas which are consist of 2010,2011 and 2012.

I'm going to use 2010 and 2011 as training data and 2012 as forecasting.

This is an example input.

 % code
date            price
3-Jan-10  1271.87
4-Jan-10  1270.2
5-Jan-10  1276.56
6-Jan-10  1273.85
....
26-Dec-12  1419.83
27-Dec-12  1418.1
28-Dec-12  1402.43
31-Dec-12  1426.19

So how can I classify input, target and test data to NN ?

Answer 1

[Shashank Prasanna]

What you are looking at is timeseries modeling and not a classification network. In order to forecast a time series such as your stock data you will need an autoregressive network which will capture the trend in the stock data.

Since you have a single time series with which you want to train a network and then forecast for future, a NARX network should work best for you. I would start by pointing you to this example in the MATLAB documentation that does exactly that. The GUI should guide you through it and generate code for future use or manipulation:

http://www.mathworks.com/help/nnet/gs/time-series-prediction.html

When you run ntstool, choose the NAR network since you don't have an external influence (such as another stock etc).

[a]

I have tried to use NARX with nnstart.

I'm still confuse to classify the input. I'm going to use 2010 and 2011 as training data and 2012 as forecasting.

date            price
3-Jan-10  1271.87
4-Jan-10  1270.2
5-Jan-10  1276.56
6-Jan-10  1273.85
....
26-Dec-12  1419.83
27-Dec-12  1418.1
28-Dec-12  1402.43
31-Dec-12  1426.19

So what is my input, target and test data look like?

[Shashank Prasanna]

If you choose NAR network, there are no inputs! there isn't even an option for you to provide one, did you try to run "ntstool" in matlab and choose NAR network?

provide your 2010 and 2011 data combined in one big vector, without the time data, just the price data only. You can choose any percentage of data for testing and validation. I would leave the default.

I urge you to run the ntstool and try it first, let me know if you have questions or if you are still unable to run it.

[a]

Thank you for you help, I got 1 step ahead forecasting with NAR network.

Then I need 5 step ahead forecasting, what should I do?

[Greg Heath]

Change the 1 to 1:5

Greg

[a]

Where?

[Shashank Prasanna]

In the generated code you can use the closed loop network for multi-step prediction.