EViews 14: stochastic simulations with linked vs unlinked equations

[vm1066]

Hi,

I had questions regarding stochastic simulations with bootstrapped innovations in EViews 14.

Firstly, on the model set-up:
I am thinking of experimenting with two versions of loading equations into the model.
1. If I use linked equation objects in the model, they are automatically loaded in as stochastic objects with specific standard deviation
2. If I use the equation representation which provides substituted coefficients and load in into the model as text, it loads in the equation as stochastic by default but standard deviation is initially set to NA. I can then use model.innov to specify standard deviations for each equation.
If I ensure standard deviations are set to same value across both the above versions of the model, will I get the same results while performing stochastic simulations?

Secondly, on the stochastic simulation options:
I wanted to see if it is possible to select the equations from which innovations are drawn. For instance, I would like to see simulation results for the entire system by incorporating uncertainty only in one equation. Can I use model.innov in this respect (perhaps by setting a stochastic eqn to be an identity instead)?

Finally, with regards to simulation output:
1. I get a message saying 'x residuals per equation calculated for bootstrap'. Is there a way to see/graph these residuals? On what basis is the number of residuals calculated?
2. Once simulation is complete, is it possible to extract values for a specific endogenous variable for all its 1000 simulations?

Thanks in advance for the help.

[EViews Matt]

Hello,

Plenty to answer here...

1. If I use linked equation objects in the model, they are automatically loaded in as stochastic objects with specific standard deviation
2. If I use the equation representation which provides substituted coefficients and load in into the model as text, it loads in the equation as stochastic by default but standard deviation is initially set to NA. I can then use model.innov to specify standard deviations for each equation.
If I ensure standard deviations are set to same value across both the above versions of the model, will I get the same results while performing stochastic simulations?

Yes, but only when using the default "Normal random numbers" innovation generation method. When using bootstrapped innovations, the equation innovation variances are ignored.

I wanted to see if it is possible to select the equations from which innovations are drawn. For instance, I would like to see simulation results for the entire system by incorporating uncertainty only in one equation. Can I use model.innov in this respect (perhaps by setting a stochastic eqn to be an identity instead)?

Yes. You can control which equations include uncertainty by flagging them as identities or not.

1. I get a message saying 'x residuals per equation calculated for bootstrap'. Is there a way to see/graph these residuals? On what basis is the number of residuals calculated?

EViews doesn't provide any direct access to the pool of bootstrapped innovations. However, the pool will just be the equations' residuals taken from a specific sample. That bootstrap sample is either:

1. The explicit bootstrap sample provided as part of the solve specification (on the "Stoch. Options" dialog page or via the 's' option to the 'stochastic' procedure), or...
2. All observations preceding the solution sample (which may default to the workfile sample) so long as there are at least 10 such observations, otherwise...
3. All observations.

2. Once simulation is complete, is it possible to extract values for a specific endogenous variable for all its 1000 simulations?

You can specify that the solve operation should record data for all stochastic variables' repetitions to a separate workfile page. From there you can examine any variable of interest.

[vm1066]

Hi Matt,

Thanks for the quick response—it has been very helpful.

I'm interested in using bootstrapped innovations because my model contains a mix of linear and nonlinear equations. Based on your explanation, I do not need to load the equations as linked objects to perform stochastic simulations i.e. I should be able to work directly with the text-based equations without specifying standard deviations.

I wanted to confirm my understanding that, if I rerun the stochastic simulation, the only reason the results would differ is that EViews draws a different set of bootstrapped innovations each time the program is executed. For the sake of reproducibility, is there a way to set a random seed (or an equivalent option) so that the simulation produces identical results each time I run the program?

Thanks again for your help.

[EViews Matt]

Hello,

Your understanding appears to be correct. You can use the 'rndseed' command to set the internal state of EViews' random number generators and create reproducible simulations. As an aside, be aware that the order in which equations are specified in the model may still interact with the fixed random number sequence to produce different results, so you cannot freely permute the list of equations and expect identical results. Reproducibility is contingent on an identical specification (that criterion is actually overly strict, but it's simple to state).