Chapter 35 Learning to Forecast Rationally

This chapter reports several robustness checks. OJ specifications designed to capture prior beliefs and non-linear responses to news detected some transient effects in many subjects, but for the most part the final regression is indistinguishable from the basic specification presented in this chapter . Eight parameter MD specifications that allow separate learning for each level of each symptom also converged roughly to the basic specification presented above, but we detect a general bias towards overresponding to the more informative symptom levels and underresponding to the less informative symptom levels. This chapter draws three conclusions from the data analysis. First, the rationality assumption is a good first approximation to subjects' forecasts at the end of 480 learning trials. Second, systematic biases towards under or overresponse can be detected in specific circumstances, e.g., overresponse in the noisier OJ environment. Third, more experiments are needed in a wider variety of tasks and environments in order to understand more fully when people can learn to forecast rationally.