Stephan Heblich

I have just posted my survey paper “Deep Learning for Solving Economic Models” on my webpage: sas.upenn.edu/~jesusfv/Deep_… In one or two weeks, it will also circulate as a working paper at the NBER and CEPR. Still, I wanted to let people know already, since I am quite happy with the outcome, largely thanks to some fantastic early feedback I got. As I have often argued, the ongoing revolution in deep learning is transforming how we solve dynamic equilibrium economic models. At its core, solving a model amounts to approximating unknown target functions (such as the value function of agents, a decision rule, or a best response function). Deep learning frequently does a fantastic job at that task. In the paper, I emphasize that this success is not “magic,” but rather the direct consequence of deep learning’s ability to discover better representations of the relevant variables of a model (for example, the state variables). The layers of a neural network transform the input variables into informationally efficient representations that can be more easily approximated. Tom Sargent loves to say that finding the state is an art. Deep learning tries to automatize that art as much as possible. This is why, in many cases, we can now solve high-dimensional problems that were computationally infeasible only a few years ago. Furthermore, the structure of deep networks designed for solving these models, largely linear apart from the non-linearity encapsulated in the activation function, permits massive parallelization. The survey paper is designed to start from the ground up. My intended audience is a first-year graduate student with only a very basic knowledge of solution methods, or even a motivated senior undergraduate. I would very much appreciate feedback. Can you follow the arguments throughout? Are there steps that remain unclear? I have taught courses based on this material at Penn, the Bank of Spain, Cambridge, the ECB, Harvard, Johns Hopkins, Northwestern, Oxford, Princeton, UC Santa Barbara, and Stanford, but I am always looking for fresh eyes to suggest improvements. All the slide decks, with links to the code, are available here: sas.upenn.edu/~jesusfv/teach… under “Machine Learning for Economists.” Eventually, I may use this survey paper and the slide decks as the kernel for something longer, but first, I need to clear my desk of too many ongoing projects.