Rugarch non convergent forecasts.
Brian thanks for reply, I?m leaning towards your suggestion by expanding window length for non converged samples. And here I got some more specific questions about rugarch package. Does anybody know is there a way in rugarch to resume ugarchroll class with new window parameters, it seems that I can resume with different solver parameters, but not with new specification. Second question is there some prebuilt method to return forecast with non converged samples filled by NAs, cause by default if non converged samples present the forecast info with vars and density don?t returns at all.
On Oct 1, 2015, at 19:37, Brian G. Peterson <brian at braverock.com> wrote:
On Thu, 2015-10-01 at 19:29 +0300, Evgeny Laba wrote:
I?m doing some VaR backtesting with garch modeling applied to stocks using rugarch package, my backtesting period is 10 years and my moving.window length is 252 & refitting parameter is one, so unsurprisingly that setup and forecast length result in some few non converged samples in some series (no more than 100), so my question is what is the best practice, rule of thumb etc., in case if I want to fill these gaps??? Changing refitting parameter to higher number, say 30 or 50 eliminate non convergence, but of cause the forecasting result is significantly different in that case?
I would generally suggest a larger rolling window, or an expanding window, for your estimates. Windowing effects can be quite severe on daily data, so for volatility forecasts I would tend towards larger data sets rather than smaller ones. Regards, Brian -- Brian G. Peterson http://braverock.com/brian/ Ph: 773-459-4973 IM: bgpbraverock