Thread Subject: Optimal regression model

First of all I need to say that I'm new with matlab enviroment. I have been trying to find a solution to my problem for several days with no results.

The problem is about finding a optimal regression model. The Y value has to be described by six variables:

Y= alfa(1)X1+alfa(2)X2+alfa(3)X3+alfa(4)X4+alfa(5)X5+alfa(6)X6

Every variable can be transformed in many ways like: 1/X1, X1^(1/2), X1^(1/3)..., so there more than 1000 combinations of the model. Its very time consuming to make it manually, test models one by one. I need matlab to test all the combinations and to find best fitting model to my Y data.

I will aprreciate any tips to my problem.

Gresha wrote:
> First of all I need to say that I'm new with matlab enviroment. I
> have been trying to find a solution to my problem for several days
> with no results.
>
> The problem is about finding a optimal regression model. The Y value
> has to be described by six variables:
>
> Y= alfa(1)X1+alfa(2)X2+alfa(3)X3+alfa(4)X4+alfa(5)X5+alfa(6)X6
>
> Every variable can be transformed in many ways like: 1/X1, X1^(1/2),
> X1^(1/3)..., so there more than 1000 combinations of the model. Its
> very time consuming to make it manually, test models one by one. I
> need matlab to test all the combinations and to find best fitting
> model to my Y data.
>
> I will aprreciate any tips to my problem.

I don't know if there is a toolbox w/ prepackaged facilities or not; you
might check the TMW web site if purchasing same isn't out of the
question. Also, of course, check the file exchange for any suitable
user-contributed tools.

After that, if one were to itemize the models to be tested, one could
write a script to automate most of the "grunt work" of computing the models.

--

On Nov 26, 8:16 am, "Gresha " <mi...@wp.pl> wrote:
> First of all I need to say that I'm new with matlab enviroment. I have been trying to find a solution to my problem for several days with no results.
>
> The problem is about finding a optimal regression model. The Y value has to be described by six variables:
>
> Y= alfa(1)X1+alfa(2)X2+alfa(3)X3+alfa(4)X4+alfa(5)X5+alfa(6)X6
>
> Every variable can be transformed in many ways like: 1/X1, X1^(1/2), X1^(1/3)..., so there more than 1000 combinations of the model. Its very time consuming to make it manually, test models one by one. I need matlab to test all the combinations and to find best fitting model to my Y data.
>
> I will aprreciate any tips to my problem.

In other words you are looking for 6 weights and functions ai,fi(.),
i=1:6
to optimize the model

Y= SUM(i=1:6){ a(i)*fi(Xi) }

I have the feeling that you are barking up the wrong tree.

1. Why must the model be of this form?
2. Why no constant term?
3. Why no interaction terms containing more than one variable?
4. What is wrong with the standard models like polynomials
and neural networks?

Hope this helps.

Greg

 
> In other words you are looking for 6 weights and functions ai,fi(.),
> i=1:6
> to optimize the model
>
> Y= SUM(i=1:6){ a(i)*fi(Xi) }
>
> I have the feeling that you are barking up the wrong tree.
>
> 1. Why must the model be of this form?
> 2. Why no constant term?
> 3. Why no interaction terms containing more than one variable?
> 4. What is wrong with the standard models like polynomials
> and neural networks?
>
> Hope this helps.
>
> Greg

This model should look like this, its used to forecast economic capital in bank. Of course there is a constant term, I forgot to put it in my equation. But my problem is not about the model. I am looking for a script, tollbox or anything which help me in finding optimal model.

Anyway thanks for response

On Nov 27, 9:48 am, "Gresha " <mi...@wp.pl> wrote:
> > In other words you are looking for 6 weights and functions ai,fi(.),
> > i=1:6
> > to optimize the model
>
> > Y=  SUM(i=1:6){ a(i)*fi(Xi) }
>
> > I have the feeling that you are barking up the wrong tree.
>
> > 1. Why must the model be of this form?
> > 2. Why no constant term?
> > 3. Why no interaction terms containing more than one variable?
> > 4. What is wrong with the standard models like polynomials
> > and neural networks?
>
> > Hope this helps.
>
> > Greg
>
> This model should look like this, its used to forecast economic capital in bank. Of course there is a constant term, I forgot to put it in my equation. But my problem is not about the model. I am looking for a script, tollbox or anything which help me in finding optimal model.
>
> Anyway thanks for response-

After you have made your variable transformations the
system of equations can be reduced to the form

Y = alpha*X

with the solution

alpha = Y/X

Hope this helsp.

Greg