Simulated method of moments targeting correlations

Dear forum,

I would like to ask principally about how to address targeting correlations when using the built-in “method_of_moments” function with “mom_method = SMM.”

As outlined in the manual, in the “matched_moments” block only linear product moments are supported, i.e., xx corresponds to the variance E[x^2] or xy corresponds to the covariance E[x*y]. However, I would like to target correlations. I have tried targeting the variance of x and of y and then their covariance to indirectly target the correlation of x and y, but that did not bring me close to the targets at all. I have also thought about introducing some auxiliary variables defined as normalized variables by their standard deviations, but I was not successful.

michal_example_smm.txt (8.7 KB)

The background is that I have my own SMM routine based on “fmincon” which iterates over Dynare simulations and which was bringing me quite close to the moments. I wanted to use the built-in Dynare routines to improve upon the result, but it does not work, presumably because I cannot deal with the correlation directly.

Would you have some suggestions or proposals please?

Many thanks in advance for your time and help!

Cheers,

Michal

No, that is not (yet) supported. The reason is the routine for computing the optimal weighting matrix, which handles only product moments, not correlations.

Thank you, Johannes, for the prompt reply. I see. Yet still, would you have please some suggestion/trick/workaround to go for correlations? Before I change the targets themselves, I would like to experiment with the correlations directly.

Thank you once again!

If you target covariance and variances, wouldn’t you be targeting correlation indirectly?

Implicitly, yes. But the weighting is harder to control to actually target the correlation.

1 Like

Thank you for replying to my post. It helped me rethink the strategy.