By Elliot Aurissergues
In this paper, I argue that agents may prefer learning a misspecified model instead of learning the rational expectation model. I consider an economy with two types of agent. Fundamentalists learn a model where endogenous variables depend on relevant exogenous variables whereas followers learn a model where endogenous variables are function of their lagged values. A Fundamentalist is like a DSGE econometrician and a follower is like a VAR econometrician. If followers (resp. fundamentalists) give more accurate forecasts, a fraction of fundamentalists (resp. followers) switch to the follower model. I apply this algorithm in a linear model. Results are mixed for rational expectations. Followers may dominate in the long run when there are strategic complementarities and high persistence of exogenous variables. When additional issues are introduced, like structural breaks or unobservable exogenous variable, followers can have a significant edge on fundamentalists. I apply the algorithm in three economic models a cobweb model, an asset price model and a simple macroeconomic model.
This horse race is a bit different from the ones focusing on the forecasting ability of statistical and micro-founded models. Here it is about how well agents following each strategy do in a fictitious world. It turns out “fundamentalists” do not do too well when the model changes on them. The critical aspect here is whether they are really blind to what is happening here. Says for example that suddenly a government loses the ability to borrow. A real-world fundamentalist would be able to reevaluate with this information. In this paper, though, he continues with a model that is obviously misspecified and only over time realizes that there is a new constraint. What is the more likely scenario?