I suppose that the main scope of an econometric models should be predictive or causal inference. Following this perspective it was shown that underspecified model can perform better than the correct specified (see here: When will a less true model predict better than a truer model?). Note that in this vein the so called “true model” is to be considered like a structural/causal model.
Now, It is known that models like ARMA are used for pure prediction and do not have a role in causal inference (see here: https://en.wikipedia.org/wiki/Autoregressive%E2%80%93moving-average_model#Estimating_coefficients); Indeed sometimes authors speak about of “free of theory models”. Moreover we can argue that endogeneity in forecasting do not have a role, and the precise value of parameters is not important; Indeed only the minimization of measure like $MSE$ is (see here: Endogeneity in forecasting). Indeed, regardless of the specification, parameters of predictive regression always maintain a some correlational meaning but this is usually of little interest.
However the tool “true model” is traditionally used in ARMA framework too. Then, considering what said above, we can show that, for example, if the $AR(2)$ is the true model we can achieve a better predictive performance from an estimated $AR(1)$ regression (underspecified) than an $AR(2)$ (rightly specified).
Now, what role the "true" $AR(2)$ model have? More precisely, what meaning/legitimacy have his two parameters? This question can be immediately generalized to any ARMA model.