The RATS Software Forum

HI, Dear all:

i am programming rolling window using maximise (say, the size of window is 500), but it is hard to guarantee the convergence for 500 moving dataset with the same initial value. It is possible to manually adjust initial value for every dataset.
I tried to replace the initial value with the forecasted coefficient from the last loop, but it does not make any better. Is there any way to find an appropriate set of initial value that ensure convergence for all loop ?
Thanks for any help!

What type of model are you estimating? I'd be very concerned if feeding previous estimates through didn't work reliably. If you're using BFGS, you might try also using the HESSIAN=%XX option. Starting BFGS from scratch when you're possibly already near the optimum might not work well.

TomDoan wrote:What type of model are you estimating? I'd be very concerned if feeding previous estimates through didn't work reliably. If you're using BFGS, you might try also using the HESSIAN=%XX option. Starting BFGS from scratch when you're possibly already near the optimum might not work well.

Thanks, Tom
i just tried, it shown that the HESSIAN expects a 13*13 matrix, but receive 1*1 matrix instead, and i realized that i should sex an initial coefficient matrix for the first loop. I tried again, it shown the same problem. I am confused.

TomDoan wrote:What type of model are you estimating? I'd be very concerned if feeding previous estimates through didn't work reliably. If you're using BFGS, you might try also using the HESSIAN=%XX option. Starting BFGS from scratch when you're possibly already near the optimum might not work well.

btw, i am doing rolling GARCH forecasting, using Maximise

You'll have to post the code. If %XX isn't dimensioned the same as your MAXIMIZE, you must be doing a secondary regression inside the loop.