Modelling Learning as Modelling
Where structural change itself is an issue, this must be reflected in continual updates of the underlying formulation of agent's models in order to reduce their forecasting bias. Indeed, one likely reason for agents to change their forecasting models is the presence of systematic bias in their forecasts.
In this paper, we describe a framework to study how agents might reduce their forecasting bias by changing the structural forms of their respective models of their environments. That is, we model economic agents as model-building forecasters rather than simply model-using forecasters who know only a given structural model.
The models our agents specify result from their observations of data and the success with which their models forecast variables of interest to them. In an environment in which there is no structural change, it also seems reasonable to suppose that, once an unbiased model with tolerable precision has been found, the process by which agents develop their forecasting models is not especially interesting. There are, however, cases in which agents' forecasts are biased or imprecise to such a degree that they would reasonably seek to improve their models. In effect, agents (like economists) might devote resources to the improvement of their forecasting models. Such improvements in forecasting capabilities would be especially important during times of structural change as, for example, in emerging market economies, where there is technological change, or perhaps at the turning point of a trade cycle.
Thus the models of learning we propose below should be particularly applicable to situations of structural change, where by "structural change" we mean that it is not merely a case of changing the parameters of the "correct" model but also changing the form of that model. A second reason for using the type of model described below, is that the agent models are meaningful. You can trace the reasons that an agent develops a particular model and hence takes a particular action. In traditional terms this is a simulation model rather than a black-box model. This is important if your primary aim is to examine and understand a process, as opposed to calculating a prediction of a future state.
In order to avoid necessary confusion we will refer to our model of how agents learn as simply models and the models built by the agents themselves as agent models or, when the agent context is clear, internal models.
Generated with CERN WebMaker