Learning to Detect Change

Abstract

People, across a wide range of personal and professional domains, need to detect change accurately. Previous research has documented systematic shortcomings in doing so, in particular, a pattern of over- and under-reaction to indications of change, resulting from a tendency to overweight signals of change at the expense of the environment that produces the signals. This investigation considers whether this pattern persists when participants are given the opportunity to learn. We find that the pattern of system neglect does persist, but that the impact of experience varies greatly across environments -- participants show reliable improvement in some conditions and virtually none in others. We explain this differential learning by formally characterizing environments in terms of the extent to which they: (i) provide consistent feedback; and (ii) tolerate non-optimal behavior. Whereas we find that learning is related to consistent feedback, the stronger -- and perhaps more surprising -- finding is that more learning occurs in environments that are more tolerant of non-optimal behavior.

Keywords: change-point detection, regime shifts, learning, overreaction, underreaction, Bayesian updating

JEL Classification: C91, D83