While I appreciate the attempt to make macroeconomics more computational, I believe DSGE goes about it in the wrong direction. In an ideal world, I'd like to see models like the Leontief Input/Output model come back to fruition. In Leontief's model (which is often given as an example in many undergrad linear algebra classes), the economy is divided into many sectors. Data is organized on each sector to estimate its influence on other sectors. In an age where data is so vast, I just don't understand how one can decide that building deeper macro theories is a good idea. We need better empirical models, not better theoretical models (we have enough of those).
This is a little inaccurate, the purpose of these macroeconomic models is either to make future predictions or run simulations to see what happens when exogenous shocks occur. They're not "testing data against theory", the data is used for parameter estimation and then verifying the accuracy of the models. It's actually very similar to the way certain AI models are developed and trained.
I do agree that these models are usually pretty inaccurate and somewhat useless though.
As an intellectual exercise to show that (for example) menu costs or nominal wage rigidities or technology shocks alone could have the economy's accounted for a shift from a [purely theoretical] equilibrium over a time period they're very interesting. For identifying which aspect contributed most towards economic change they offer little, and as a policymaking tool they're actually worse than useless since the models are generally built on the economists' assumptions about agents' response to policy and fitted to the data to justify those assumptions, rather than using the data to understand how responsive, rational and optimizing agents actually are. As such, their forecasting performance isn't very good either...
http://economics.sas.upenn.edu/~schorf/papers/bayesian_macro...
(this is the closest I could find to a self-contained link, sorry). This stuff's not great, but I don't know what you mean by "useless for prediction". And, given a model, it's pretty trivial to figure out which shocks contributed when; this is the whole point of Impulse Response Functions, variance decompositions, etc.
I've said this elsewhere on the page, but I'll repeat it here: this stuff isn't popular because it's logically airtight and compelling, but because it seemed to have done pretty well empirically. So the criticism that dsge models are unrealistic isn't very interesting; everyone knows that already, especially the people who use the stuff (for the most part. I'm sure there are some dsge truthers too). If anyone has an approach that works better empirically at addressing core macro questions, especially the newly important interplay between the real economy and credit markets, this would be a great time to put it out there. There are a lot of people paying attention.
Here's my understanding (I am not an expert): DSGE is designed to give an approximate response to exogenous shocks for NL models. For that, it needs to start from an equilibrium.
You seem to say that shocking from an equilibrium is silly. Can you explain me why? I wonder what alternative solution you have in mind that could improve DSGE.