As a data scientist at the New York Fed, my experience with economists and their willingness to try 'new' methods has been mixed, but generally positive. We've managed to do some very interesting projects and work using deep learning, especially in the NLP space.
What I've always been surprised about is that I don't know of any economist that's tried to do any macro-economic modeling using agent based reinforcement learning.
What economists who practice it mean by “agent based economic modeling” is a nothingburger with no influence on the wider field of economics.
Generally what we economists want is to solve an inverse problem:
1. Suppose that agents are behaving according to a utility function with a set of parameters, let’s estimate those parameters. (Utility functions are restricted in form by economic theory.)
(One Example: an individual’s utility function may be characterized by a degree of risk aversion - given observed data on many individuals, can we recover their risk aversion?
Another Example: a firm’s investment decisions may depend on (intrinsically unobservable) fixed costs - given observed data on firm behavior in some industry, can we recover fixed costs which are compatible with the model and the observed decisions?)
2. Given the parameters we estimated, what sort of interesting counterfactual policies can we study, under the assumption that the parameters we estimated above don’t change?
This is why you see “less reinforcement learning than you may expect”. Sure, I can write down a model of a macroeconomic environment and put some reinforcement learning agents in it who may learn to behave optimally given the environment. Are the parameters governing those RL models relevant to individuals and their decisions? I have no idea.
Compare the alternative (inverse) approach of writing down a plausible reinforcement learning model and figuring out what parameters govern agents’ decisions from observed decision data. It’s a very different problem.
An interesting perspective on agents in human factor complex situations of scientific analysis as part of the overall progression of the scientific process since Galileo is here :
Like I said in another comment, I'm not an economist. My background is in CS and though I work with economists at times when they need the help, my skills are often more useful in operations.
I do enjoy it! Well I got into it here specifically mostly by coincidence, but a lot of our team started as interns while doing their masters and have a background in quant finance, stats, or CS. I do post on the HN Who's Hiring threads, so you can keep an eye out there as well.
I think the application of AI in economics and finance is going to have muffled utility over time if not already because of feedback. In other words certain results will be expected based on historical data, however the future will greatly differ from the past because of AI intervention which will skew expected results. In turn that will make AI essentially useless in an economic/financial application.
“AI” for prediction from historical trends is not very interesting in my opinion. Much more interesting is complexity economics with agent based simulation.
Presumably, what is meant is a prediction non based on patterns, but on a simulation of economic dynamics made through the interaction of not deterministically coded virtual agents within some environment.
Is it just me or is economics (or at least these authors) trying to shoehorn its way into being something it's not by saying AI has effected economic behavior?
This is the same profession that literally pays the Nobel Prize organization to give them a special medal every year.
If you want to understand the condition of economics as a study, simply search for 'is Economics a Science?'. Economics will do anything that will give them legitimacy that their research can not provide.
What I've always been surprised about is that I don't know of any economist that's tried to do any macro-economic modeling using agent based reinforcement learning.