I guess Paul and some others were looking for an excel garch(1,1) tool. I´ve found this like a year ago at Harvey Campbell´s page (in case I dont upload it right, maybe it is still there with some interesting papers and videos on asset management where Harvey shows he is a gentleman: http//www.duke.edu/~charvey.I hope it is usefull and helps with my promotion one day...(I´ve compared the results with eviews and it works most of the time).Now a question: I have been reading about weather derivatives and I`ve learned they use models like Vasicek or Cir to model rain and temperature because they are mean reverting. I would like to know, if possible, how do I get the mean revertion quoficient from a historic series...some kind of optimization? Thanks a lot! (dont tell me same as with interest rate models cause I dont know how to deal with them neither!) Thanks again

where Harvey shows he is a gentleman >>I've never met Cam, but we have about 5 mutual friends, all of whom say he is one of the nicest, classiest people around. I would like to know, if possible, how do I get the mean revertion quoficient from a historic series...some kind of optimization? Thanks a lot! (dont tell me same as with interest rate models cause I dont know how to deal with them neither!) Thanks again >>I have attached a little text file showing how to do it in Excel. (I hate trying to write equations in the forums -- variable pitch fonts bad, monospaced good...) If you need more detail, I'll add it. (Its a .zipped text file--the forum won't let me upload straight text...)HTH,matt.

Thanks for your help matthew. I see you are in Ohio...what can you tell about Rene Schultz? Iggy

Mahoffer,pretty much all that I can tell you is that his name is Rene Stulz. I've only been at OSU for 3 months, and my path doesn't seem likely to cross Rene's too often. Matt.

I wonder about estimation of parameters in GARCH(1,1).Hull (Options, 4th ed, pp. 376) says, "A general purpose algorithm such as Solver in Microsoft's Excel is liable to provide a local rather than global maximum of the likelihood function."When I tried to estimate GARCH(1,1) using approach described by Hull with Solver I really failed to.However using Harvey Campbell´s spreadsheet and Solver it is possible to estimate GARCH(1,1). Estimated parameters are almost equal to those obtained with SAS (a powerful statistical package). I have slightly modified Harvey's procedure to estimate process y(t)~N(mu,sigma), where sigma is GARCH(1,1). But in this case Solver sometimes estimates parameters that are obviously a local maksimum and so a good guess is needed to obtain a global maksimum. Nevertheless Excel help states, "Microsoft Excel Solver uses the Generalized Reduced Gradient (GRG2) nonlinear optimization code developed by Leon Lasdon and Allan Waren". To my knowledge, this method designed to search for a global maksimum.Can anybody say what the problem is with using Solver to estimate GARCH(1,1)? Does anybody have a reliable C/C++ code to estimate GARCH(1,1)?

Spartak,In reality, Spartak, the issue of global vs. local maximization is quite complicated. In my experience, vanilla GARCH(1,1) processes are typically concave, and therefore the local optimum = global optimum. The Excel solver should be adequate for standard GARCH(1,1). Another issue of GARCH estimation is scale; the typical dependent variable in GARCH estimation are price returns, i.e. (p(t)-p(t-1))/p(t-1) or ln(p(t))-ln(p(t-1)), in either of these cases, it will greatly improve your results to estimate 100*return instead of return directly. (Sorry that this is short, I'm on vacation for the holidays. I tend to be more verbose when I'm at the office )If you'd like an application to compare your excel results against, in another thread I posted a small .exe app that I wrote to perform GARCH estimation. (look under the thread "Little GARCH tool..." under the student forum.) But frankly, if you have access to SAS, I'd use it, those guys know what they're doing when it comes to statistical estimation.HTH,Matt.

matthewcroberts,I have a problem running your GARCH.exe. I enter commands as you prescribed but I can't see the results of operations since after execution there appear some words and the programm quickly closes. I think the problem is caused by my Windows98. Do you know how to cope with it?<< it will greatly improve your results to estimate 100*return instead of return directly >> When I use it I get that mean (mu) increases in 100 times and arch0 term increases in 10000 times while arch1 and garch1 terms remain the same with comparison to unscaled time series. Should I make adjustment after estimation or can it be done running estimation?Concerning SAS I see its major drawback (at least in V8.1) that it is not possible to make forecasts using GARCH(1,1), which is needed in many practical applications. As a matter of fact, I initially estimate GARCH(1,1) and then use my procedure written in IML module to get forecasts with estimated parameters. It doesn't seem rational to me.Happy New Year!

spartak,try running the garch.exe program from inside a command-line window. That will allow you to see the error (which I'd appreciate if you'd mail to me at roberts.628<at>osu.edu)As for the scaling of the data, note the scaling should only be performed if you are using returns series. For simulation, scale the simulated series, not the parameters. Also, for reference, what are the params of your GARCH process before & after scaling (or at least their magnitudes?)As for SAS, I didn't know that it doesn't have simulation (I use Matlab now.) but that is really odd.Matt.

matthewcroberts,I managed to run your GARCH.exe.I used it to estimate one univariate time series and I found that the estimated parameters a rather different from those estimated with SAS (mainly, b (arch1) and w (arch0)).I don't understand what is mean covariate and variance covariate data in your program? Could you explain it?

Spartak,because of the work I do, I am often interested in other things that might influence variance. Therefore, my program was written to estimate the model: y(t) = m + x(t)'*c + e(t) e(t) = sqrt(h(t))*u(t) h(t) = w + z(t)'*d + a*h(t-1) + b*e(t-1)^2 whereu(t) ~ N(0,1)the covariates in the mean are the x(t); i.e. they allow the value of the observed series to be a function of other observable variables, say, day of week effects, the covariates in the variance are the z(t) which allow you to make variance a function of other observable data, such as the season, or for event studies. If the data is not proprietary, I'd appreciate it if you could send it to me. I can't vouch too much for this code, as I wrote & tested it 2 years ago, but at that time, I deemed it sufficiently useful. Perhaps you have uncovered a problem that I didn't find. The estimates should be quite similar.Matt.