On the plausibility of adaptive learning in macroeconomics: A puzzling conflict in the choice of the representative algorithm

The literature on bounded rationality and learning in macroeconomics has often used recursive algorithms such as least squares and stochastic gradient to depict the evolution of agents' beliefs over time. In this work, we try to assess the plausibility of such practice from an empirical perspective, by comparing forecasts obtained from these algorithms with survey data. In particular, we show that the relative performance of the two algorithms in terms of forecast errors depends on the variable being forecasted, and we argue that rational agents would therefore use different algorithms when forecasting different variables. By using survey data, then, we show that agents instead always behave as least squares learners, irrespective of the variable being forecasted. We thus conclude that such findings point to a puzzling conflict between rational and actual behaviour when it comes to expectations formation.