Markov Switching VECM

[nacrointfin]

Dear All:

Can the markov switching vecm be implemented under the command 'msvarsetup'?

Best,

Terence Wan

[TomDoan]

Dear All:

Can the markov switching vecm be implemented under the command 'msvarsetup'?

Best,

Terence Wan

You would have to use @MSSysRegression instead since it allows for more general equation forms. The one difference is that MSVARSETUP allows for switching means, while MSSysRegression only has switching coefficients.

[nacrointfin]

Dear Tom

Is there any markov switching VECM example using command @MSSysRegression ?

Best,

Terence Wan

[TomDoan]

This has some extra calculations, but does shows the general idea. You estimate the cointegrating vector by standard methods then input the error correction term as an explanatory variable.

Code: Select all

open data data.xls
calendar(q) 1955
all 2006:01
data(format=xls,org=columns) 1955:01 2006:01 cgbv cgce bcaj dydc royj bcje
*
* Patch over missing values in BCAJ, which is quarterly for part of the range.
*
dec rect a(2,2)
compute a=||2.0,-1.0|1.0,0.0||
compute c=||1.0,0.0||
dlm(type=smooth,sw=%diag(||1.0,0.0||),a=a,c=c,exact,y=bcaj) 1959:2 * xstates
set bcaj 1959:2 * = xstates(t)(1)
*
set le  = log(royj/bcaj)
set lp  = log(cgbv)
set lwp = le-lp
set lns = log(dydc/(1-.01*bcje))
set ly  = log(cgce)
set ln  = log(dydc)
*
graph
# ly 1965:1 2001:1
graph
# lwp 1965:1 2001:1
*
source johmle.src
@johmle(determ=rtrend,lags=4,eigval=eigval) 1966:2 2001:1
# ly lwp ln lns
*
compute [symm] s10_00_01=tr(%%s01)*inv(%%s00)*%%s01
dec rect d(5,3)
input d
   0.0    1.0    0.0
   0.0    0.0    1.0
   1.0    0.0    0.0
  -1.0    0.0    0.0
   0.0  -.0056 -.0048
eigen(general=%mqform(%%s11,d)) %mqform(s10_00_01,d) eigvalr eigvecr
*
* LR test of restriction given 1 cointegrating relation
*
compute lr=%nobs*log(1-eigvalr(1))-%nobs*log(1-eigval(1))
cdf chisqr lr 2
*
compute [vect] x1r=d*%xcol(eigvecr,1)
compute x1r=x1r/x1r(4)
disp "Restricted cointegrating vector"
disp x1r
*
* Error correction series
*
set eqm = %dot(x1r,||ly,lwp,ln,lns,t||)
diff(center) eqm 1965:2 2001:1
*
set dly  = 100.0*(ly-ly{1})
set dlwp = 100.0*(lwp-lwp{1})
set dlns = 100.0*(lns-lns{1})
set dln  = 100.0*(ln-ln{1})
*
@mssysregression(states=2,switch=ch,nfix=8)
# dly dlwp dlns dln
# dly{1 2} dlwp{1 2} dlns{1 2} dln{1 2} constant eqm
*
compute gstart=1965:4,gend=2001:1
*
nonlin(parmset=msparms)  p
nonlin(parmset=vecmparms) sigmav gamma beta
@MSSysRegInitial gstart gend
*
* Reset beta(2) to zero loadings on the cointegrating vector
*
compute beta(2)(2,1)=beta(2)(2,2)=beta(2)(2,3)=beta(2)(2,4)=0.0
*compute sigmav(2)=sigmav(1)*2.0
@MSSysRegEMGeneralSetup
do emits=1,50
   @MSSysRegEMStep gstart gend
   disp "Iteration" emits "Log likelihood" %logl
end do emits
compute p=%xsubmat(p,1,nstates-1,1,nstates)
frml logl = f=%MSSysRegFVec(t),pt_t1=%mcstate(p,pstar),$
  pt_t=pstar=%msupdate(f,pt_t1,fpt),log(fpt)
maximize(start=(pstar=%MSSysRegInit()),parmset=vecmparms,$
  method=bfgs) logl gstart gend