It's funny in a way, that a modern person believes that through technology we can very well reach consciousness with computers, go to Mars, design our genes, achieve immortality, colonize the universe, find the "god's particle", and so on, but successfully modelling an economy is impossible...
It's not impossible (I would like to believe) but it involves self-referential models: In order to work, economic models must include models of the models (both yours and other peoples'.) This means that cybernetics (the study of "circular loops in causation") is the minimum formalism for economics.
Consciousness was referenced above. One thing commonly associated with consciousness, at least in humans, is a theory of mind. We model the actors all around us on the assumption that they each have their own mind, and that by modeling their mind we can predict their actions and intentions given enough situational context.
Mind you, theory of mind might not be a prerequisite for consciousness and maybe it's even possible to have theory of mind without consciousness, but the point is that machines acting in environments that are affected by other actors are going to develop models for those actors naturally as they're pushed to become more proficient in their role. Keep pushing them to become more proficient at anything involving multiple actors and it's plausible that the machine's model for a given actor will itself contain predictions for that actor's models of the other actors -- i.e. self-referential models. That Go AI could probably predict the opponent's action multiple turns down the road, which includes predicting its own response to each move, for example.
I think the more significant difference between modeling a Go opponent and modeling an economy is not self-referential models, but probably just the number and depth of connections. In the progression of training machines to play games, we've gotten good at single-player games like Mario, then two-player games like Chess or Go, and now some work is going into multiplayer games like Starcraft. Each of these represents a slightly "larger" closed environment.
Contrast this to a frictionless economy, where each actor's actions have sweeping effects across the globe. Usually tiny effects, but they reach far through the interconnections. Unlike these games, the approach of starting by modeling a small, closed economy and then scaling that up to larger economies isn't nearly as feasible. There really aren't any "independent" economies today to study. The best you could do is point to a mostly-isolated state like North Korea, but even that's a huge thing to model on its own and the data for doing that isn't accessible. I think the bigger issue with modeling economies isn't the self-referential models -- those are relevant in other areas that machines are already becoming proficient at -- but the scale and the fact that any on-ramp to help deal with that scale incrementally is not obvious.