Do humans process data like Stata? An experimental study

Abstract

Least square (LS) learning model is one of the most seminal models on how individuals can learn a rational expectation equilibrium (REE) if they do not initially start from there. According to this model, agents estimate the data generating process (DGP) of the market price using the ordinary least square (OLS) model in an iterated way. In this paper, we test whether and how agents converge to REE in the lab, and replace the prediction task in the Learning to Forecast Experiment (LtFE) from point prediction to parameters in the DGP. About 17% of the individual predictions can be categorised to follow the LS learning rule, though there is a lack of evidence indicating the adoption at the aggregate level. We also design two treatments to investigate the effect of the spread of the independent variable on the speed of learning. Our results show that the speed of learning and the occurrence of convergence is much higher when the spread of the independent variable (“weather”) of the DGP is larger. In accordance with econometric theory, we also find a smaller variance in the treatment with wider spread using an experimental approach, though dispersion between the two treatments is not statistically significant.