Regime-Switching Stochastic Volatility and Short-term Interest Rates,
by Mdhu Kalimipalli and Raul Susmel
Appendix A: The Gibbs Algorithm for Estimating the RSV Model
In the RSV model (2), we need to estimate the parameter vector θ ={β, γ, ση, φ1 ,
p01, p10 } along with the two latent variables Ht = {h1 ,...,ht } and St ={s1 ,….,st }. Thus, the
parameter set consists of ω = {Ht , St , θ} for all t. We use Bayes theorem to decompose
the joint posterior density as follows.
f ( H n , S n ,θ ) ∝ f (Yn H n ) f ( H n S n ,θ ) f ( S n θ ) f (θ )
We next draw the marginals f(Ht | Yt , St ,θ), f(St |Yt ,Ht,θ), and f(θ|Yt , Ht St ), using
the Gibbs sampling algorithm described below:
Step 1:
Specify initial values θ(0) ={β 1(0) , γ(0) , σ η,(0) ,φ (0), p01 (0), p10 (0) }. Set i =1.
Step 2:
Draw the underlying volatility using the multi- move simulation sampler described in De
Jong and Shephard (1995), based on parameter values from step 1. The multi- move
simulation sampler is used to draw the underlying volatility vector for all the data points
as a single block (see De Jong and Shephard (1995) for details). Consider the RSV
model (3), reproduced below:
∆ rt − (aˆ 0 + aˆ 1 rt −1 ) ≡ RES t
RES t = ht rt 2−α1 ε t , α = 0 .5
(ln (h ) − µ ) = φ (ln( h
t st 1 t −1 )
) − µ s t − 1 + σ η2 η t −1
µ s t = β + γs t γ >0 s t = {1 ,2} (3)
The conditional mean equation can be written as,
ln( RES t2 ) = ln( ht ) + ln( rt −1 ) + ln( ε t2 ) (A - 1)
The term ln(ε t 2 ) can be approximated by a mixture of seven normal variates (Chib,
Shephard, and Kim (1998)).
ln (et2 ) = z t
( )
7
f(z t ) = ∑ f N z i mi − 1.2704,v i2 i = { 1,2 ,....7 } (A - 2)
i =1
Now, (A-1) can be written as
ln( RES t2 ) = ln( ht ) + ln( rt −1 ) + [z t k t = i ] (A - 3)
where kt is one of the seven underlying densities that generates zt . Once the underlying
densities kt , for all t, are known, (A-3) becomes a deterministic linear equation and, along
with the RSV model (3), can be represented in a linear state space model. Next, apply the
De Jong and Shephard (1995) simulation smoother to extract the underlying log volatility
from the observed data.
In order to estimate α as a free parameter, rewrite (A-1 ) as
ln( RES t2 ) = ln( ht ) + 2α ln( rt −1 ) + ln( ε t2 ) (A - 1)′
Then estimate α, approximating ln(ε t2 ) by a lognormal distribution. Once α is known,
follow (A-3) and extract the latent volatility.
Step 3:
Based on the output from steps 1 and 2, the underlying kt in (A-3) is sampled from a
normal distribution as follows -see Chib, Shephard and Kim (1998):
[ ] (
f z t= i ln( yt2 ), ln( ht ) ∝ q i f N z i ln( ht ) +m i − 1.2704, vi2 ) i≤k (A - 4)
For every observation t, we draw the normal density from each of the seven normal
distributions {k t = 1,2,..,7}. Then, we select a “k” based on draws from uniform
distribution.
Step 4:
Based on the output from steps 1, 2 and 3, we draw the underlying Markov-state
following Carter and Kohn (1994). We use the smoother for the above state-space model
(3), to derive the vector of underlying state variable St , t = 1,2,...,n.
Step 5:
Cycle through the conditionals of parameter vector θ ={β, γ, ση, φ1 , p01, p10 } for the
volatility equation using Chib (1993), based on the output from steps 1-4. Assuming that f
(θ) can be decomposed as:
f (θ Yn ,H n , S n ) ∝ f ( β Yn , H n , S n ,θ − β ) f (γ Yn , H n , S n ,θ −γ ) f (σ 2 Yn , H n , S n , θ −σ 2 )
f (φ Yn , H n , S n , θ −φ ) f ( p01 , p10 Yn , H n , S n , θ − pij ) (A - 5)
where θ-j refers to the θ parameters excluding the jth parameter. The respective
conditional distributions (normal for β, γ and φ, inverse gamma for σ2 and beta for pij ) are
described in Chib (1993). The parameter γ is drawn using an inverse CDF with the
restriction that it is positive. The prior means and standard deviations are specified in
Tables 3 and 5.
Step 6: Go to step 2.
Estimation of SSV model (2) has the same steps as in RSV model (3), except that we do
not have to draw the latent states and transition probabilities. For the Gibbs estimation,
we leave out the first 4000 draws (i.e., burn–in iterations are 4000) and sample from the
next 6000 draws. We choose every fifth observation to minimize, and if possible
eliminate, any possible correlation in the draws. Our effective number of draws therefore
drops to 1200 (i.e., effective test iterations are 1200). We construct 95% confidence
intervals for the parameters, based on 1200 draws. We construct the standard errors for
the parameters using the batch- means method -see Chib (1993). We estimate the density
functions for the parameters using a Gaussian kernel estimator (Silverman (1986)). The
kernel estimator with kernel K is defined as:
^ 1 n x−X
f (x ) = ∑ K i
nh i = 1 h
where h is the band width, Xi {1,...…,n} are the observations from the given sample and
n is the sample size. When K is a normal density, we get the Gaussian kernel.
Appendix B: A Monte Carlo Experiment of the RSV, using the Gibbs Algorithm
We perform a Monte Carlo experiment of the RSV model (2), without level effects, i.e.,
we consider the following version of model 2:
RES t = ht ε t
(ln (h t ) ( )
) − µ s t = φ 1 ln( h t − 1 ) − µ st − 1 + σ η2 η t − 1
µ s t = β + γs t γ >0 s t = {1,2}
We first assume values for the parameter set θ ={β, γ, ση, φ 1 , p01, p10 }-we designate them
as “true” values; the true parameter values used in the simulation are listed in the table
below. Using the true transition probabilities p01 and p10 , we generate a state vector (with
values 0 or 1) of size 1000. Using the state vector and the true parameters β, γ, ση, and φ 1 ,
we generate stochastic volatility, i.e ht . Then, we generate the residual vector RES t based
on the RSV model described above. Then, taking RES t as given, we estimate the
parameter set θ using the MCMC algorithm as explained in Appendix A. We set the
number of burn–in iterations equal to 4000 and the number of effective test iterations
equal to 1200. We construct the 95% confidence intervals for the parameters based on
1200 draws. We estimate the standard errors for the parameters using the batch- means
method -see Chib (1993). The results are reported in Table B.1.
Table B.1
Results from a Monte Carlo experiment
Parameter True Prior Values Posterior Values
values
Mean Std. Deviation Mean (Std. error) Std. deviation 95% Confidence Interval
β 0.7 0 50 0.759 (0.003) 0.112 (0.500- 0.986)
γ 1.5 1 50 1.477 (0.018) 0.191 (1.051 -1.819)
φ 0.4 0 1 0.391 (0.005) 0.085 (0.202- 0.540)
σ2 0.6 - - 0.733 (0.005) 0.109 (0.538 - 0.976)
p 01 0.01 0.2 0.16 0.011 (0.000) 0.007 (0.003-0.035)
p 10 0.04 0.2 0.16 0.074 (0.003) 0.036 (0.026-0.166)
*Prior distribution of σ2 (inverse gamma) is improper. Sample size is T: 1000.
We find that the posterior means of parameters are quite close to the true values.
The standard errors are small, indicating a high precision of the posterior means. For the
variance and the transition probability p10 , the posterior means are slightly higher than
true values. However, they clearly lie within the 95% confidence bounds.
Figure B.1 shows the latent volatility and states. The top panel consists of
simulated residuals RES t obtained using the true parameter vector θ. The second panel
presents both the true and latent volatility, the latter obtained using the simulation
smoother. The latent smoother volatility appears to be smoother compared to the true
volatility. This is not surprising because the smoother volatilities are smoother by
construction compared to the filter volatilities. To investigate how close the simulated
and true ht are to each other, we calculate the Wilcon test statistic and we find that it is
0.79, i.e., insignificant at 5% level. The third panel presents the true states –i.e., either 0
or 1- and smoother probabilities of being in the high vo latility state. We see that the
smoother probabilities track the latent volatility quite well. For example, for the second
and third high volatility regimes, we find that the smoother probabilities begin to drop as
the latent volatility begins to drop.
Figure B.1. Simulated Monte-Carlo Residuals and Corresponding Latent Volatility and
States