Least squares learning? Evidence from the laboratory

We report on an experiment testing the empirical relevance of least squares (LS) learning, a common way of modelling how individuals learn a rational expectations equilibrium (REE). Subjects are endowed with the correct perceived law of motion (PLM) for a price level variable they are seeking to forecast, but do not know the true parameterization of that PLM. Instead, they must choose and can adjust the parameters of this PLM over 50 periods. Consistent with the E-stability of the REE in the model studied, 97.8% of subjects achieve weak convergence to the REE in terms of their price level predictions. However, the number of participants that can be characterized as least squares learners via the adjustments they make to the parameterization of the PLM over time depends on properties of the data generating process of the dependent and independent variables. Participants learn the REE faster, and behave more like least squares learners when there is greater variance in the independent variable of the model. We consider several alternatives to least squares learning and find evidence that many subjects employ a simple satisficing approach.