Learning to believe in simple equilibria in a complex OLG economy - evidence from the lab

We set up a laboratory experiment within the overlapping-generations model of Grandmont (1985). Under perfect foresight this model displays infinitely many equilibria: a steady state, periodic as well as chaotic equilibria. Moreover, there exists some learning theory predicting convergence to each of these equilibria. We use experimental evidence as an equilibrium selection device in this complex OLG economy, and investigate on which outcomes subjects most likely coordinate. We use two alternative experimental designs: learning-to-forecast, in which subjects predict the future price of the good, and learning-to-optimize, in which subjects make savings decision. We find that coordination on a steady state or 2-cycle are the only outcomes in this complex environment. In the learning-to-forecast design, coordination on a 2-cycle occurs frequently, even in the chaotic parameter range. Simulations of a behavioral heuristic switching model result in initial coordination on a simple AR(1) rule though sample autocorrelation learning, with subsequent coordination on a simple second-order adaptive rule once the up-and-down pattern of prices has been learned.