12.07.2015 Views

"Frontmatter". In: Analysis of Financial Time Series

"Frontmatter". In: Analysis of Financial Time Series

"Frontmatter". In: Analysis of Financial Time Series

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

428 MCMC METHODSTable 10.2. Estimation <strong>of</strong> Bivariate Volatility Models for Monthly LogReturns <strong>of</strong> IBM Stock and the S&P 500 <strong>In</strong>dex from January 1962 to December1999. The Stochastic Volatility Models Are Based on the Last 1000 Iterations<strong>of</strong> a Gibbs Sampling with 1300 Total Iterations.(a) Bivariate GARCH(1, 1) model with time-varying correlationsParameter β 01 β 02 α 10 α 11 α 12 α 20 α 21 γ 0Estimate 1.04 0.79 3.16 0.83 0.10 10.59 0.04 0.35Std. Error 0.31 0.20 1.67 0.08 0.03 0.93 0.02 0.02(b) Stochastic volatility modelParameter β 01 β 02 α 10 α 11 σ 2 1vα 20 σ 2 2vγ 0 σ 2 uPost.Mean 0.86 0.84 0.52 0.86 0.08 1.81 0.39 0.39 0.08Std. Error 0.30 0.18 0.18 0.05 0.03 0.11 0.06 0.03 0.02These prior distributions are relatively noninformative. We ran the Gibbs samplingfor 1300 iterations, but discarded results <strong>of</strong> the first 300 iterations. The randomsamples <strong>of</strong> g ii,t were drawn by Griddy Gibbs with 400 grid points in the intervals[0, 1.5s 2 i ],wheres2 iis the sample variance <strong>of</strong> the log return r it . Posterior means andstandard errors <strong>of</strong> the “traditional” parameters <strong>of</strong> the bivariate stochastic volatilitymodel are given in Table 10.2(b).To check for convergence <strong>of</strong> the Gibbs sampling, we ran the procedure severaltimes with different starting values and numbers <strong>of</strong> iterations. The results are stable.For illustration, Figure 10.7 shows the scatterplots <strong>of</strong> various quantities for two differentGibbs samples. The first Gibbs sample is based on 300 + 1000 iterations, andthe second Gibbs sample is based on 500 + 3000 iterations, where M + N denotesthat the total number <strong>of</strong> Gibbs iterations is M + N, but results <strong>of</strong> the first M iterationsare discarded. The scatterplots shown are posterior means <strong>of</strong> g 11,t , g 22,t , q 21,t ,σ 22,t , σ 21,t , and the correlation ρ 21,t . The line y = x is added to each plot to showthe closeness <strong>of</strong> the posterior means. The stability <strong>of</strong> the Gibbs sampling results isclearly seen.It is informative to compare the GARCH model with time-varying correlationsin Eqs. (10.33)–(10.36) with the stochastic volatility model. First, as expected, themean equations <strong>of</strong> the two models are essentially identical. Second, Figure 10.8shows the time plots <strong>of</strong> fitted volatilities for IBM stock return. The upper panelis for the GARCH model, and the lower panel shows the posterior mean <strong>of</strong> thestochastic volatility model. The two models show similar volatility characteristics;they exhibit volatility clusterings and indicate an increasing trend in volatility. However,the GARCH model produces higher peak volatility values. Third, Figure 10.9shows the time plots <strong>of</strong> fitted volatilities for the S&P 500 index return. The GARCHmodel produces an extra volatility peak around 1993. This additional peak does notappear in the univariate analysis shown in Figure 10.5. It seems that for this particularinstance the bivariate GARCH model produces a spurious volatility peak. This spuriouspeak is induced by its dependence on IBM returns and does not appear in thestochastic volatility model. <strong>In</strong>deed, the fitted volatilities <strong>of</strong> S&P 500 index return by

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!