Forecasting Behavior in Instable Environments

Summary In many economic situations periodically occurring changes in behavior of the involved agents can be observed. These changes have the characteristics of abrupt structural breaks. The behavior often seems to switch between regimes as if there were constant relationships between economic variables between these breaks. Examples are the alternating price determination of sellers and buyers on a market out of equilibrium or the periodical development of price cartels and the resulting switches in prices. In this study, we want to analyze the expectation formation of participants of a laboratory experiment subject to regime switches. Despite the practical relevance there is hardly any experimental contribution regarding the reaction of economic decision-makers to structural breaks. The only systematic experiment in this context was performed by [2]. The authors found inconsistent evidence on the performance of judgmental forecasts versus statistical procedures in the literature and therefore wanted to experimentally test the individual performance of subjects. The noise levels of the time series, the type of break (abrupt or creeping) and the direction of the break was systematically varied between 10 time series. The participants were told that the time series may contain structural changes. The judgmental forecasts were found to perform significantly worse than statistical procedures. The subjects were trying to read too much signal into the series and their forecasts contained excessive noise. In our experiment the forecasting performance will be analyzed but the main interest is the explanation of the average forecasts in order to understand how the subjects react on the break. The most simple casefor a structural break in our data generating process is a constant shift. Three time series are applied in the experiment subject to one break and two breaks respectively. Between the breaks the data generating process remains constant. By these means the behavior of the subjects in several regimes and their reactions on the breaks can be observed. Recently [1] presented a simple heuristic for the modeling of average forecasts of the subjects. The authors showed that the model forecasts the behavior of the subjects better than the Rational Expectations Hypothesis (REH) when indicators are in the information set of the participants. This heuristic is restricted to stable and stationary time series. We will apply a modified version of the model to the forecasts in the setting with the structural breaks. We find that the behavior of the subjects after the break is best described by a transition phase. When the new level of the series has established several periods after the break the information before the break (i.e. especially earlier turning points) is gradually ignored. We also find that the human performance compared to statistical procedures is rather poor.