How to fix the sum of coefficient to 1 (OLS)

[Aktar]

Hello,

i perform Frankel and wei (1994) regression and i want to fix the sum of coefficient to 1 with eviews (for interpreting coefficients loke a percentage), stangerly when i perform normal OLS i find such coefficient extremly high ... i don't know how it is possible.

ex: i regress exchange rate indonésia/australian dollar against exchange rate US/australian dollar, and i find a coefficient regression near of 9800 !

Thanks for help in advance

[EViews Glenn]

http://forums.eviews.com/viewtopic.php? ... ction#p103

[Aktar]

tnak you, so i must build my "own programm or is it possible to do this directly with estimate ?

Thank you

[trubador]

http://forums.eviews.com/viewtopic.php?f=5&t=562

[Aktar]

thank you very much

I estimated this equation

Code: Select all

c = 0 equation eq1.ls idr = (exp(c(2))/(1+exp(c(2))) * (usd - eur)) + (exp(c(3))/(1+exp(c(3))) * (amu - eur)) + (exp(c(4))/(1+exp(c(4))) * (eur - eur)) vector(4) test = c vector(4) test_norm for !i = 2 to 4 test_norm(!i) = @exp(test(!i))/(1+@exp(test(!i))) next vector test2 = test_norm/@sum(test_norm) ' coefficients sum to 1

and i find this conclusion : WARNING: Singular covariance - coefficients are not unique
i join my workfile.

Please what should i do for imporve my result, thank for your help.

i join my workfile

[startz]

You have a typo at the end where you've written eur-eur

[Aktar]

ok thank you, i tried with this equation :

Code: Select all

equation eq1.ls idr = (exp(c(1))/(1+exp(c(1))) * (usd - idr)) + (exp(c(2))/(1+exp(c(2))) * (amu - idr)) + (exp(c(3))/(1+exp(c(3))) * (eur - idr))

But it doesn't work, the estimate run but the prob. is 1 for all coefficient and these are equal to -36 (i join the new workfile). i do not understnd excatly how it work. Which variable i much soustract ? it is the explain variable (idr) ? Thank you very much for your valuable help

[trubador]

First of all, the correct form of the coefficients in your equation would be as follows:

Code: Select all

idr = (exp(c(1))/(exp(c(1)+exp(c(2))+exp(c(3)))) * (usd - idr)) + (exp(c(2))/(exp(c(1)+exp(c(2))+exp(c(3)))) * (amu - idr)) + (exp(c(3))/(exp(c(1)+exp(c(2))+exp(c(3)))) * (eur - idr))

This equation becomes nonlinear in coefficients, which makes the estimation harder. You may need to supply good initial values to obtain feasible results.
However, I think your model is ill-defined. There is a significant scale problem between your dependent and independent variables. Therefore, I think the variables should be defined in (log)returns, not the levels. Please check with your reference paper...