Gareth, 1. For such a small dataset (242 points) try using QML (i.e. distribution="norm"). 2. Also, don't use "coredata" and don't multiply by 100. Instead, pass the returns as xts (as is).
-Alexios On 17/09/2014 12:33, Gareth McEwan wrote: > Hi Alexios > > I seem to be getting exceptionally big t-values in a lot of my fitting > output (across a number of financial variables). The majority of the > variables are in "monthly log return" format calculated from "monthly > price observations" over the last 20 years (of the 18 log return series > 1 is first-differenced and another is second-differenced). Sample sizes > are small, unfortunately, only around 242 log return observations per > variable. > > I'm a big fan of reproducible research, but I'm not sure how to get the > data to you. I've attached a .csv document to this email though. If > others are interested, I've selected the FSPI variable (S&P 500 Index) > for which data is easily obtainable. My data is from a local data vendor > here in South Africa. So, this variable is in monthly log returns, > multiplied by 100 to put it in percentage form. > > # 5: S&P 500 Index (Developed Equity Markets) > # 5: ARMA(2,3)-GARCH(1,1) errors~jsu > spec.FSPI <- > ugarchspec(variance.model=list(model="sGARCH",garchOrder=c(1,1), > > submodel=NULL,external.regressors=NULL,variance.targeting=F), > > mean.model=list(armaOrder=c(2,3),include.mean=T,external.regressors=NULL), > distribution.model="jsu") > garch.FSPI <- > ugarchfit(spec=spec.FSPI,data=coredata(FSPI.log.ret),solver="hybrid") > show(garch.FSPI) > > Optimal Parameters > ------------------------------------ > Estimate Std. Error t value Pr(>|t|) > mu 0.70104 0.000762 919.8719 0.000000 > ar1 -1.93142 0.000391 -4944.8270 0.000000 > ar2 -0.93236 0.000206 -4534.2598 0.000000 > ma1 1.85979 0.000165 11287.6463 0.000000 > ma2 0.72827 0.000149 4887.0466 0.000000 > ma3 -0.13019 0.000058 -2232.2089 0.000000 > omega 1.28003 0.599599 2.1348 0.032776 > alpha1 0.21330 0.038001 5.6129 0.000000 > beta1 0.74371 0.035146 21.1605 0.000000 > skew -2.35018 1.132755 -2.0747 0.038010 > shape 2.19635 0.714951 3.0720 0.002126 > > Robust Standard Errors: > Estimate Std. Error t value Pr(>|t|) > mu 0.70104 0.029683 23.61746 0.00000 > ar1 -1.93142 0.030470 -63.38848 0.00000 > ar2 -0.93236 0.012092 -77.10479 0.00000 > ma1 1.85979 0.001194 1558.23243 0.00000 > ma2 0.72827 0.000222 3283.16473 0.00000 > ma3 -0.13019 0.003171 -41.05573 0.00000 > omega 1.28003 4.319409 0.29634 0.76697 > alpha1 0.21330 1.281223 0.16648 0.86778 > beta1 0.74371 1.042388 0.71347 0.47556 > skew -2.35018 4.813643 -0.48823 0.62538 > shape 2.19635 3.490002 0.62933 0.52914 > > The excessively large t-values are worrying me. Have you found this to > be normal? Or is there something I'm missing in the modelling methodology? > > Also, a word on how the Robust Standard Errors are calculated would be > highly appreciated. I find some of the fitted variables have > statistically significant variables in the "Optimal Parameters" output, > but which then become insignificant in the "Robust Standard Errors" > output (as in the case above). Can you provide any guidance on this finding? > > Many many thanks for the help !! > Gareth > > > > _______________________________________________ > [email protected] mailing list > https://stat.ethz.ch/mailman/listinfo/r-sig-finance > -- Subscriber-posting only. If you want to post, subscribe first. > -- Also note that this is not the r-help list where general R questions > should go. > _______________________________________________ [email protected] mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-finance -- Subscriber-posting only. If you want to post, subscribe first. -- Also note that this is not the r-help list where general R questions should go.
