EViews.com

Hello everyone,

I have a base model consisting of 62 equations, which I solved using the *stochastic options* with a diagonal covariance matrix and scaled variances to match the equation-specified innovation standard deviations. The model works correctly in this configuration.

Subsequently, I added 6 new equations to the base model. These new equations directly or indirectly affect only 3 of the original 62 equations. I verified that everything else in the model remained unchanged and that the innovations used for the existing equations are exactly the same. However, I noticed that the *solve* results for the previous equations, which are not affected by the new additions, still differ.

My question is: why do the predictions for some of the original equations change, even though they are not directly or indirectly influenced by the new equations?

Thanks a lot!!

Hard to say without having the workfile/model.

I'm not authorized to send it, could this difference be due to the "Broyden" solver?

Hello,

Given the situation you describe, I'd consider it very unlikely that the solution changes you've observed are a consequence of the underlying solver used, e.g. Broyden vs Newton. While changing the number of equations in the model will affect the sequence of innovations used across the stochastic repetitions, the end result shouldn't be significantly different (unless the model is numerically unstable and/or extremely sensitive to initial conditions). Plus, such an issue would show up by simply re-solving the base model multiple times, and that's not the situation you describe. The most likely culprit is an unrecognized dependence among the model variables which is allowing the new variables to affect a broader range of the original variables. Out of curiosity, if you perform a deterministic solve instead do you still observe a solution difference between the original model and the augmented model?

As Gareth mentioned, it's hard to diagnose the cause of this issue (model setup vs EViews bug) without the workfile.

Thank you so much for your insight—it’s greatly appreciated! I've performed a deterministic solve: I found identical solutions for the equations shared between the original and augmented models.
Furthermore, the differences are minimal and not significant, with the sole exception of one variable whose equation is entirely independent of the additions to the model. I find it strange that this happens...but, as you mentioned, it could be due to the sequence of innovations used across the stochastic repetitions.

Thanks again!!