3
$\begingroup$

I'm reading about this approach of using GARCH-EVT-copula methodology to separate univariate and joint estimation and then estimate for example VaR and ES. I wanted to try something similar, but my attempt failed miserably. Anyone of you sees a mistake in the methodology outlined below? Any hint is appreciated!

I took a few series, computed log returns, and multiplied them by -1 (so that I have losses on the right and returns on the left). Then I used this approach:

1) ARMA-GARCH fitting according to AIC (using the rugarch package);

2) obtain standardized residuals $z_t = \frac{\epsilon_t}{\sigma_t}$;

3) semi-parametric modeling of $z_t$ using a generalized Pareto distribution for upper and lower tails (thresholds at 10% and 90%) and Gaussian kernel for the interior part (using the spd package);

4) transformation in uniform margins using the pspd command from the previous package;

5) copula fitting using an Archimedean copula model via maximum likelihood (since ML could not be evaluated in all points, I transformed the uniform margins in pseudo-observations using the pobs command before fitting);

6) since I want a time-varying copula approach, I use a moving window of 1000 observations to estimate a time-varying Archimedean copula parameter, and fit an ARIMA in order to do forecasting;

7) given predicted copula parameters (thanks to previous ARIMA specification), N copula realizations are simulated from day $t+1$ to $T$;

8) using the qspd command, I transform copula realizations back in standardized residuals $z_{t}$;

9) I insert estimated $z_t$ back in ARMA-GARCH specification, compute log returns, transform in simple returns, and calculate return of equally weighted portfolio;

10) now I have N returns for each day, and can compute required VaR as the $\alpha$-quantile (90, 95, 97.5, 99, ...).

Now the staggering surprise: this methodology failed miserably for me for relatively moderate quantiles (90, 95 and 97.5) and more extreme quantiles (99, 99.5, 99.9). I obtain a number of violations which is WAY more than expected. Any reason why? Is there a mistake in the steps mentioned?

Improve this question
$\endgroup$
6
  • $\begingroup$ Are you using actual historical asset returns or simulated returns for this? $\endgroup$
    – John M
    Jun 8, 2016 at 2:53
  • $\begingroup$ I'm using 2000 historical daily log returns to estimate the ARMA-GARCH coefficients, the GPD parameters and the time series of the copula parameter. Once I have this time series, I estimate the ARIMA so that I forecast the 1-step ahead parameter and simulate N copula realizations for (t+1); then again forecasting 1-step ahead parameter and simulation of N realizations for (t+2) and so on until the end of the forecasting period. Then steps 8 to 10. $\endgroup$
    – Kondo
    Jun 8, 2016 at 7:27
  • $\begingroup$ I suggest you use a Monte Carlo simulated returns with a normal distribution to help you debug your problem. $\endgroup$
    – John M
    Jun 8, 2016 at 12:21
  • $\begingroup$ John, may I ask you what you mean exactly? Simulate random draws from a multivariate normal and interpret these as returns? Also how would I calibrate mean vector and covariance matrix if that's the case, with sample mean and covariance? $\endgroup$
    – Kondo
    Jun 9, 2016 at 13:47
  • 1
    $\begingroup$ Hi Kondo! Did you find the fault eventually? $\endgroup$
    – Richard Hardy
    Mar 6, 2019 at 16:59

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.