options double crt nwidth=25 signif=16;
name bgfcp 'TSP 4.4 GARCH(1,1) benchmark from Bollerslev-Ghysels JBES 4/96';
?
?  Below we reproduce p.146 Table 2 column 4 - GARCH(1,1) results
?          B-G       TSP 4.4 1998        FCP (VSGARCOV)       Consensus SEs
?  mu    -.006     -.0061904083799     -.006190408379994
?        (.009)    (.00918935396086)   (.009189353979760952)  (.0091893540)
?  alpha0 .0108     .0107613978518      .0107613978518
?                  (.0064931860821)    (.006493186182105218)  (.006493186)
?  alpha1 .153      .15313406182        .153134061820
?        (.054)    (.05353170253)      (.05353170349543810)   (.053531703)
?  beta1  .806      .80597367031        .805973670305
?        (.073)    (.072461448212)     (.07246144960234878)   (.07246145)
?  LogL  -1106.6  -1106.60788104129   -1106.607881041288
?--------------------
?  Notes:
?  Initial conditions are the new TSP 4.4 1998 new defaults:
?    h(0) = e(0)**2 = SSR(mu_current_iteration)/T
?  All SEs are robust (QMLE - White)
?  The convergence tolerance was set extremely low, to obtain the
?    highest possible accuracy, and the above numbers are the 11+ digits
?    which agree from two different sets of starting values.
?  y = Deutschmark-British Pound daily percentage nominal returns,
?      1/3/1984 - 12/31/1991
?      Obtained on 1/30/98 from   ftp  web.amstat.org
?      jbes/View/96-2-APR/bollerslev_ghysels/bollerslev.sec41.dat
?  Requires TSP 4.4 dated 5/11/98 or later, which has the new
?      ARCH options HINIT= , E2INIT= , HITER=N, and HCOV=W .
?  by Clint Cummins
smpl 1,1974;
read(file='dmbp.dat') y monday;
arch(nar=1,nma=1,tol=1e-15) y c;

? Now try with old 6-digit FCP results as starting values.
?   (these were from before making the 2 changes which made the FCP
?    program accurate to double precision).
? Final results from these starting values agree with the previous
?   arch command (using default starting values) to the 11-12
?   digits given in the benchmark listed above.
mmake @start
-.006190408667240
 .010761422088305
 .153134435686848
 .805973258082795 ;
arch(nar=1,nma=1,tol=1e-15) y c;

? The FCP program is from Fiorentini, Calzolari, and Panattoni,
?   Journal of Applied Econometrics, 1996, pp.399-417.  It was
?   downloaded from the JAE web page on 7/20/1998:
?   http://qed.queensu.ca/jae/1996-v11.4/fiorentini-calzolari-panattoni/
?   and the source code itself is dated 1-31-96.  It was compiled
?   with Lahey LF90 v3.51, with the -o0 switch.  It did not work with
?   Lahey F77L-EM/32 v5.01, because of adjustable array dimension errors,
?   and other problems not investigated.
?   2 changes were made to both VSGARCMX and VSGARCOV to make it accurate
?   to double precision:
?   1. Declare VALUNC as double precision (add REAL*8):
?      REAL*8 FUNCTION VALUNC(IDYNAM,C, ...
?   2. Extend LOG(SQRT(2*PI)) in VALUNC to double precision:
?      .9189385 --> .9189385332046725D0
?      (this makes the Log Likelihood accurate)
?   A few other changes were added to print the LogL and Standard Errors
?   to 16 digits.  The FCP results for the table above were from a single
?   set of starting values.
? The command below can be used to write out the data to 8 digits
?   for use by the FCP program.
?write(file='bg.2',format='(5G15.8)') y;