PARCH

[puf]

Hi, thank you for fixing previous problems. However, now I have a major one. GARCH(1,1) model in eviews sometiems predicts negative volatility. This ovviously does not make any sense. I believe that the problem is in the optimisation routine, which probably does not inclode neccesary restrictions. In case of GARCH(p,q) models, all the estimated constants should be larger or equal then zero. Not only that this is standard in al the existing literature, but it will definitelly improve forecasting performance of the GARCH models in Eviews. Here is the code to illustrate my point:

Code: Select all

wfopen 1.csv
scalar length = @obsrange
series logr = dlog(abs(prc)/cfacpr)
series forecasted_garch
scalar estimation_window = 100
for !i=estimation_window to length-2
      smpl @first+!i-estimation_window @first+!i
      equation kamo.ARCH(1,1) logr
      smpl @first+!i+1 @first+!i+1 
      kamo.forecast r se var
      forecasted_garch(!i+2)=var(!i+2)
next
smpl @all
genr problem = @recode(forecasted_garch<0,1,0)
scalar nr_of_neg_vol_forecasts = @sum(problem)

As you can seen, in this case GARCH(1,1) predicts negative volatility in 10 cases. When you open the equation "kamo", you can see that one of the estimated coefficients is negative.
Could you please incorporate coefficient restrictions in the GARCH estimations? Preferably not just for GARCH(p,q) models where all the estimated coefficients should be nonnegative, but for any other models where these restrictions should apply.

[EViews Gareth]

Unfortunately EViews does not perform constrained optimisation. Further, in the case of GARCH, it is usually the case that negative coefficients are a useful indicator of a badly specified model.

However we take your point, and will investigate the possibility of improving the routines in the future.