COMPGROUPS.NET | Search | Post Question | Groups | Stream | About | Register

### Financial forecasting

• Email
• Follow

```I am a final year under graduate. I am trying to build a system to
predict stock market indices. As an output, my desirable output only
consists of 1,0,-1 but after training the network the outputs are not
discrete but a continous variable. Why is it? I am using a network with
2 hidden layers. So theoretically it should model disrete variables. So
how should i interpret the results? Or should i change the network?

Thanks

```
 0
Reply bsayanthan (2) 7/27/2005 4:34:18 PM

See related articles to this posting

```Saya wrote:
> I am a final year under graduate. I am trying to build a system to
> predict stock market indices. As an output, my desirable output only
> consists of 1,0,-1 but after training the network the outputs are not
> discrete but a continous variable. Why is it? I am using a network with
> 2 hidden layers. So theoretically it should model disrete variables. So
> how should i interpret the results? Or should i change the network?

Use 3 binary {0,1} targets and logsig output activation functions.
The continuous outputs can be interpreted as posterior probabilities
conditional on the input. Choose the output with the largest value.
If you want to save the probability estimates, scale the outputs
so that the scaled outputs sum to 1.

Hope this helps.

Greg

```
 0

```
Greg Heath wrote:
> Saya wrote:
> > I am a final year under graduate. I am trying to build a system to
> > predict stock market indices. As an output, my desirable output only
> > consists of 1,0,-1 but after training the network the outputs are not
> > discrete but a continous variable. Why is it? I am using a network with
> > 2 hidden layers. So theoretically it should model disrete variables. So
> > how should i interpret the results? Or should i change the network?
>
> Use 3 binary {0,1} targets and logsig output activation functions.
> The continuous outputs can be interpreted as posterior probabilities
> conditional on the input. Choose the output with the largest value.
> If you want to save the probability estimates, scale the outputs
> so that the scaled outputs sum to 1.

Now that it's a little later in th AM, I think I should have asked
what the 3 values stand for. Buy/NoAction/Sell ? Then stay with
one logsig output train to {0,1/2,1} targets and use a validation
set to find Buy/Sell thresholds. That way, you still have the
probability interpretation.

Hope this helps.

Greg

```
 0

```On 27-Jul-2005, "Saya" <bsayanthan@gmail.com> wrote:

> I am a final year under graduate. I am trying to build a system to
> predict stock market indices. As an output, my desirable output only
> consists of 1,0,-1 but after training the network the outputs are not
> discrete but a continous variable. Why is it? I am using a network with
> 2 hidden layers. So theoretically it should model disrete variables. So
> how should i interpret the results? Or should i change the network?

I am the author of a decision tree program called DTREG.  I would like to
run your data through DTREG and compare the results with those that you are
getting from your neural network.  Would it be possible for you to e-mail
your data file to me along with a description of the variables
(target/predictors, categorical/continuous)?  I will run it through DTREG
and send you the results so that we can compare how a decision tree based
model (TreeBoost and Decision Tree Forests) compares with the neural network
model.

--
Phil Sherrod
(phil.sherrod 'at' sandh.com)
http://www.dtreg.com  (decision tree modeling)
http://www.nlreg.com  (nonlinear regression)
http://www.LogRover.com (Web statistics analysis)
```
 0

```Thanks Every one
Indeed i got a continous function but with all the values near those
discrete values after altering my architecture to have a tanh layer as
the output layer. I was expecting perfect matching. And it was to
signify buy ,sell and no action.

Saya wrote:
> I am a final year under graduate. I am trying to build a system to
> predict stock market indices. As an output, my desirable output only
> consists of 1,0,-1 but after training the network the outputs are not
> discrete but a continous variable. Why is it? I am using a network with
> 2 hidden layers. So theoretically it should model disrete variables. So
> how should i interpret the results? Or should i change the network?
>
> Thanks

```
 0

```Oh wow i would like to see this comparison.  Did this materialize?  Is
there some url with the comparisons?

Thanks.

```
 0

5 Replies
72 Views

Similar Articles

12/8/2013 10:23:11 PM
[PageSpeed]

Similar Artilces:

forecasting 135098
hi folks, i want to do forecasting with backpropagation algorithm. I wrote the code for backpropagation. but I am confused with the input and desired response. suppose i have one month 10 minutes interval wind speed data. i want to predict.my question in this case what should be the desired output? the algorithm output should be predicted output, right? please help me regarding this as I am new in neural network. thanks.looking forward to reply your mail. ashraf n.b:FYI i use matlab 7.0 On 17-Aug-2008, ashraf.pata@gmail.com wrote: > i want to do forecasting with backpropagation algorithm. I wrote the > code for backpropagation. but I am confused with the input and desired > response. suppose i have one month 10 minutes interval wind speed > data. i want to predict.my question in this case what should be the > desired output? the algorithm output should be predicted output, > right? For time series forecasting, the input usually consists of lagged values of the value you are trying to predict. For example, let's say you are trying to predict hourly power consumption for a utility which we will call 'P'. Then you would create a training file

dcc forecasting
Hi, i am doing some research on the correlations between commodity values. I used the uscd-garch toolbox to compute the parameters for the dynamical correlations model but it lacks a tool for forecasting future correalations.... if somebody has a matlab code for forecasting with dcc model or know where to get it please post it here. thx Regards Navid

Forecasting Forex with NN
My question, I think, is a simple one. But first some background. I am attempting to do the following for a class. I am wanting to build a NN that will predict (with any degree of accuracy, for now) the closing price for the next day, based on the previous X days. I have obtained some historical data. I am using JOONE to model and design the NN. My basic question is, what type of data should I be giving the NN? Are raw closing prices O.K.? Do I need to manipulate the data first? I could also use the help of someone that has any experience with JOONE. Thanks, chansen On Sep 20, 6:20 pm,

Re: Forecasting monthly sales by product #2
Fi, Generally, you need to look at either the High-Performance Forecasting solution or SAS/ETS (perhaps using Enterprise Guide as a front-end). Do you have access to any of these? SAS/STAT is not sufficient for forecasting. What industry are you in? Retail or ?? What factors are currently considered to influence demand? You mention 'monthly sales' and 'daily budget' - what kind of seasonal/cyclic influences affect your products? (Are daily, weekly or monthly fluctuations significant?). A bit more information would help get better responses from the 'L. HTH, Dave. -----Original Message----- From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On Behalf Of David Johnson Sent: Thursday, 12 July 2007 10:37 PM To: SAS-L@LISTSERV.UGA.EDU Subject: Re: Forecasting monthly sales by product I didn't see any replies here. Can't you classify the products by the amount of data and use that in your model? Intuitively I would expect new products may have some variation caused by leading edge and gadget buyers (surely the iPhone release is proof of that idea) so as the product matures we find it is less popular, or the market saturates and its sales