Seriously, how many free parameters can you add before this just becomes an estimation exercise on a nonlinear system. Any model with decent out of sample performance is just a black box.
Unpopular Opinion: DSGE is a fad and will be replaced by VARs

The "if it doesn't fit the data, just add more parameters" approach to DSGE modeling is an embarrassing disaster. But that's certainly not the only approach to DSGE modeling, and more sensible approaches still show a lot of promise.
And VARs? Really? They're fine for finding interesting correlations, but nothing more than that.

Welcome to the frontier of 1995, OP.
The problem is that the empirical macroeconomic research since the 90s is pretty useless for any policy maker. DSGE models only deliver results based on the priors of the modeler and don't bring in any useful information.
At least, simple VARs, although without identification, are linked with the real world and relevant correlation is still much more useful than fake structural effect.

And VARs aren’t a black box?
Reduced form regressions are a black box as well.
Not really
Ok I'm listening what is your reasoning.
VAR are just OLS: people understand what is going on, simple correlations.
DSGE are calibrated and loglinearized, the internal mechanism is completely opaque, especially far away from the longrun equilibrium

I was replying the poster who was saying reduced form regressions are not a black box. But as a reply your response, what do you mean people understand OLS, as in how it's estimated, sure but the derivations for how Bayesian estimations on DSGE models are also available (more technically demanding). I figured we'd call these methods black boxes because of the lack of deciphering the internal mechanisms. The point of DSGE is an attempt just at that.

I was replying the poster who was saying reduced form regressions are not a black box. But as a reply your response, what do you mean people understand OLS, as in how it's estimated, sure but the derivations for how Bayesian estimations on DSGE models are also available (more technically demanding). I figured we'd call these methods black boxes because of the lack of deciphering the internal mechanisms. The point of DSGE is an attempt just at that.
DSGE purposefully obfuscates economic mechanisms to allow the modeler to achieve the results he/she wants. Using calibration instead of estimation, loglinearization instead of analytical/numerical solutions is not science: this is voodoo.

DSGE purposefully obfuscates economic mechanisms to allow the modeler to achieve the results he/she wants. Using calibration instead of estimation, loglinearization instead of analytical/numerical solutions is not science: this is voodoo.
A good model makes the assumptions and mechanism clear. Loglinearization is an approximation  what's wrong with that? Calibration admittedly gives the researcher free variables, but no more than any other modeling assumption.

DSGE purposefully obfuscates economic mechanisms to allow the modeler to achieve the results he/she wants. Using calibration instead of estimation, loglinearization instead of analytical/numerical solutions is not science: this is voodoo.
A good model makes the assumptions and mechanism clear. Loglinearization is an approximation  what's wrong with that? Calibration admittedly gives the researcher free variables, but no more than any other modeling assumption.
Because most economists don't understand mathematics.
Loglinearization is basically a Taylor expansion, only valid around the neighborhood of the expanded point (often the steady state). However, economists infer dynamics from their loglinearized results far away from the steady state, which is plain stupid.
"Gives researcher free variables": then you should understand how ridiculous it is and why this is absolute garbage and unscientific. I can understand if this is done for very minor third or fourth order coefficients as it is done in physics. But calibrating risk aversion, which determines the system first order, is just wrong and dishonest.

DSGE purposefully obfuscates economic mechanisms to allow the modeler to achieve the results he/she wants. Using calibration instead of estimation, loglinearization instead of analytical/numerical solutions is not science: this is voodoo.
A good model makes the assumptions and mechanism clear. Loglinearization is an approximation  what's wrong with that? Calibration admittedly gives the researcher free variables, but no more than any other modeling assumption.
Because most economists don't understand mathematics.
Loglinearization is basically a Taylor expansion, only valid around the neighborhood of the expanded point (often the steady state). However, economists infer dynamics from their loglinearized results far away from the steady state, which is plain stupid.
"Gives researcher free variables": then you should understand how ridiculous it is and why this is absolute garbage and unscientific. I can understand if this is done for very minor third or fourth order coefficients as it is done in physics. But calibrating risk aversion, which determines the system first order, is just wrong and dishonest.Why do you think we rely on calibration and loglinearization?