On the complexity of forming mental models

We experimentally study how people form predictive models of simple data generating processes (DGPs), by showing subjects data sets and asking them to predict future outputs. We find that subjects: (i) often fail to predict in this task, indicating a failure to form a model, (ii) often cannot explicitly describe the model they have formed even when successful, and (iii) tend to be attracted to the same, simple models when multiple models fit the data. Examining a number of formal complexity metrics, we find that all three patterns are well organized by metrics suggested by Lipman (1995) and Gabaix (2014) that describe the information processing required to deploy models in prediction.