CODEWITHSUNDEEP
rcurve-fittingdistributionfgarch
smaica
smaica

Reputation: 817

How to support the convergence of student-t distribution fits (stdFit and sstdFit) from fGarch package in R?

I want to fit a student-t and a skewed student-t distribution to log returns. For this purpose I use the stdFit and sstdFit function from the fGarch package (would also use something else if someone knows a better option for both distributions).

My problem is, that the fit in most of the cases does not converge. Here some example code:

set.seed(123)
require(xts)
require(rugarch)
require(fGarch)
data(sp500ret)
spx = as.xts(sp500ret[1000:nrow(sp500ret),,drop=F])

hist(spx,200)

fit_std = stdFit(spx)
if(fit_std$convergence != 0){
  print("stdFit did not converge")
}

fit_sstd = sstdFit(spx)
if(fit_sstd$code != 0){
  print("sstdFit did not converge")
}

The only additional parameters I can pass to stdFit and sstdFit are parameters for the optimizer nlm: https://www.rdocumentation.org/packages/stats/versions/3.6.2/topics/nlm

I would rather look for something like start value and / or boundaries.

Also I did not find an explanation to what the other options of fit_sstd$code mean. I found code having the values of [0,1,2,3] at least.

Any ideas how I can improve the fitting? I need to repeat different fits automatically > 1 million times.

Thanks a lot!

Edit:

Here is my sessionInfo() output to reproduce:

R version 4.2.3 (2023-03-15 ucrt)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 10 x64 (build 19045)

Matrix products: default

locale:
[1] LC_COLLATE=German_Germany.utf8  LC_CTYPE=German_Germany.utf8    LC_MONETARY=German_Germany.utf8 LC_NUMERIC=C                   
[5] LC_TIME=German_Germany.utf8    

attached base packages:
[1] parallel  stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] fGarch_4022.89 rugarch_1.4-9  xts_0.13.1     zoo_1.8-12    

loaded via a namespace (and not attached):
 [1] tidyselect_1.2.0            remotes_2.4.2               ks_1.14.0                   reshape2_1.4.4             
 [5] lattice_0.20-45             colorspace_2.1-0            vctrs_0.6.2                 generics_0.1.3             
 [9] SkewHyperbolic_0.4-0        utf8_1.2.3                  rlang_1.1.0                 pracma_2.4.2               
[13] gbutils_0.5                 pillar_1.9.0                glue_1.6.2                  cvar_0.5                   
[17] lifecycle_1.0.3             plyr_1.8.8                  stringr_1.5.0               timeDate_4022.108          
[21] munsell_0.5.0               fBasics_4022.94             gtable_0.3.3                timeSeries_4021.105        
[25] mvtnorm_1.1-3               fansi_1.0.4                 Rcpp_1.0.10                 KernSmooth_2.23-20         
[29] scales_1.2.1                DistributionUtils_0.6-0     truncnorm_1.0-9             spd_2.0-1                  
[33] ggplot2_3.4.2               stringi_1.7.12              dplyr_1.1.1                 rbibutils_2.2.13           
[37] numDeriv_2016.8-1.1         cowplot_1.1.1               grid_4.2.3                  Rdpack_2.4                 
[41] cli_3.6.1                   tools_4.2.3                 magrittr_2.0.3              Rsolnp_1.16                
[45] tibble_3.2.1                pacman_0.5.1                spatial_7.3-16              pkgconfig_2.0.3            
[49] GeneralizedHyperbolic_0.8-4 MASS_7.3-58.2               Matrix_1.5-3                rstudioapi_0.14            
[53] R6_2.5.1                    mclust_6.0.0                compiler_4.2.3

Upvotes: 1

Views: 89

Answers (1)

IRTFM
IRTFM

Reputation: 263451

I eventually got the packages installed and my initial errors fixed. At that point I was getting the same results as you did. Looking at the code I wondered whether giving an entire xts object (rather than a "vector of quantiles" as described in the help page) could be the source of problems. So I tried:

fit_std = stdFit(spx[,1])
Warning messages:
1: In nlminb(start = start, objective = loglik, lower = c(-Inf, 0,  :
  NA/NaN function evaluation
2: In nlminb(start = start, objective = loglik, lower = c(-Inf, 0,  :
  NA/NaN function evaluation
3: In nlminb(start = start, objective = loglik, lower = c(-Inf, 0,  :
  NA/NaN function evaluation
> if(fit_std$convergence != 0){
+     print("stdFit did not converge")
+ }

So it converged. I then went back and ran it with stdFit(spx[,1]). This time it converged, although the warning persisted. So was the initial lack of convergence due to your choice of the seed? I then tried recreating the lack of convergence to allow investigation. Again it converged. So did running a successful fit change the starting values?

Opinion: Your effort to fit multiple series should proceed by passing atomic vectors extracted from the xts objects and you should continue with the checks for convergence.

Upvotes: 0

Related Questions