New Experimental Evidence on Expectations Formation

In this paper, we measure belief formation in an experimental setting where agents are incentivized to provide accurate forecasts of a random variable, drawn from a stable and simple statistical process. Using these data, we estimate an empirical model that builds on the recent literature on expectation dynamics: It nests rational expectations, but also allows for extrapolation and under-reaction. Our findings are threefold. First, the rational expectation hypothesis is strongly rejected in our setting. Second, both extrapolation and underreaction patterns are statistically discernible in the data, but extrapolation quantitatively dominates. Third, our model coefficients are very robust to changes in experimental setting: They do not depend on process parameters, individual characteristics or framing. These large and stable deviations from rationality occur even though the forecasting exercise is simple and transparent.