Generating Confidence Intervals in ARDL Model Simulations on EViews

[Bob]

Hello,

I am working with EViews 13, coding to simulate dynamic effects from an ARDL model. When I incorporate the ARDL equation into an EViews model to perform stochastic simulations similar to those I do with a VAR model, the trajectories for the median and the confidence interval bounds in the ARDL model are identical, without variation. With a VAR, I receive appropriate confidence intervals through stochastic simulations. Is there a specific configuration or option that I should adjust to obtain valid confidence intervals in ARDL simulations?

Thank you for your help!

[EViews Gareth]

Could you provide an example?

[Bob]

To be more precise, there are no differences in the scenario gaps (see the last chart from my sample program.

[EViews Matt]

Hello,

I believe you're seeing correct behavior. The only difference between the two scenarios is a one unit shock to exogenous variable X1 on date 2019m01. Given that the model is entirely linear, any difference between variable trajectories in the two scenarios will be a function of solely your explicit shock, not the random shocks simulated at every repetition. The individual trajectories do vary by repetition of course, as summarized by the Y_0[HLM] and Y_SC1[HLM] series. However, since the same sequence of random shocks is used in both scenarios within a repetition, and the model is linear, in the difference those random shocks effectively cancel out. There will be no variation between repetitions, as summarized by the Y_SC1_0[HLM] series. This is basically the same idea behind VAR IRFs, where only a single round of calculations are needed since the residuals don't matter.

[Bob]

Hello Matt,
Thank you for the clarifications. Indeed, in a linear model, when the residuals are identical across scenarios (the EViews procedure), it's normal not to see variation in the simulated gaps. Perhaps a feature to vary residuals could be added in a future update. In the meantime, I could consider trying to code a bootstrapping procedure for this purpose. IMPORTANTLY: I've noticed the option that allows for integrating statistical uncertainty on coefficients (ardl_mod.stochastic(c=t)). It's not the uncertainty on the residuals, but it's just as important, and in my case, it does yield confidence intervals.