Neural Networks and Macro

Today I will briefly talk about an interesting connection between Machine Learning and Macro by a friend and colleague, Artem Kuriksha, who is working with Neural Networks and Macroeconomics. His idea is to make the policy decision function, namely that of investment, depend on a Neural Network.

How does it work? In a usual Aiyagari dynamic problem, the agents maximize their utility function through the consumption and savings decision, solving the recursive problem over a horizon. Therefore, the savings decision, called the policy function, is a closed-form function that basically depends on the state of the economy and the current savings or assets held. There are many things that could not work, though. One of them is that agents could find it hard to learn about the stochastic process. If that was the case, then they would not choose the optimal savings rate.

Nonetheless, the emphasis of his paper is to consider the policy function in itself a Neural Network. The motivation for this is that perhaps people are not able to fully solve a recursive problem as is typically formulated. If the horizon is long enough, though, this learning algorithm delivers the actual solution, which is great. Now, if the horizon is not long enough, then different experience leads to different policy functions, which could potentially explain why some agents decide to accumulate way more than they should. This property of the model can explain how people behave.

This is a very novel approach to Macroeconomics, as far as I’m concerned. There are many questions that remain open: what does it really mean to have agents use a Neural Network for optimizing a problem? One bigger challenge is the interpretation of a Neural Network. Even if one has all the estimates of the model, they are very hard to explain, as they don’t map onto something relatable. In other words, a Neural Network is basically a fancy non-linear regression. Until there is no better understanding of how the brain works and these Neural Network models, the best we have are conjectures.

One more practical problem is how information flows, as well as what structural parameters actually govern a system. Economic growth is a rather new event in Human History. It is not clear how sustainable it will be over the course of the future. These two elements represent important challenges for the present and future economists.

2 Replies to “Neural Networks and Macro”

  1. whoah this blog is wonderful i really like reading your articles. Keep up the great paintings! You realize, a lot of people are hunting round for this info, you could help them greatly.

Leave a Reply to rama Cancel reply

Your email address will not be published.