Heuristics for the subjective prediction of times series

The practice of forecasting is dominated by judgmental techniques. Due to this practical relevance the abilities of individuals to predict time series were explored in numerous experimental studies. Researchers want to find out which circumstances influence or improve the human forecasting performance and when statistical models are superior. Despite of this increasing interest in describing expectation formation there are hardly any mathematical models explaining forecasting behaviour. The bounds & likelihood heuristics, BECKER/LEOPOLD-WILDBURGER (1996), is a simple, but very effective procedure that models collective forecasts of individuals very well. There are no heuristics or applications of neural networks in literature that perform as well in modeling the expectation formation of subjects, in particular for short time series. This model was successfully tested in different experimental settings. Therefore, our intention is to generalize our promising results. The following questions are of particular interest: Can the model be applied to non-stationary time series? Can the model also explain empirically observed forecasts of experts? Can our results be applied to time series as they occur on capital markets? Is there comparable way to model individual behaviour? What are the effects of additional information (using exogenous time series) on forecasting behaviour? Can the heuristics also be applied to time series with structural breaks? Our objective is to find answers to these questions by analysing our sample of available data and conducting new experiments.

Traditionally economics is a formal science which tries to explain human behavior using mathematical models. We work in the area of experimental economics. In this branch of research the real economic decisions of subjects are observed and analyzed in a laboratory. The participants in such experiments are financially remunerated depending on their decisions. Thereby it is ensured that the obtained insights can be transferred to real situations outside of the laboratory. In our research project we analyze the decision-making behavior of subjects who forecast time series. A lot of decision-relevant information is presented in the form of time series and published in every newspaper: share prices, unemployment figures, the development of interest and inflation rates or the oil price. The forecasts of economic agents are of great theoretical interest. The expectations of future developments influence, for instance, a person`s investments in the capital market, demand for commodities or savings behavior. The participants in our experiments predict time series and are paid depending on their forecasting accuracy. There are mathematical models which describe, explain and forecast expectation formation processes. We have tested these models on the forecasts of our subjects and developed our own model (called the bounds & likelihood heuristic), which makes it much easier to carry out forecasts. Our work focused on an area which has been neglected by research so far. In reality, several sources of information are available to decision-makers. To take up the example of the investor above: He bases his decision not only on one source of information but on many: besides the share price, for instance, he also takes into consideration the state of the whole market. In our experiment we presented the subjects with several leading series as additional sources of information for the forecast of a time series. The characteristics of this time series were varied systematically between several versions of the experiment. Our model forecasts the average behavior of the subjects better than comparable models in all versions of the experiment. We have started to address the forecast of individual behavior. This is significantly more difficult because the behavior of individuals is very instable and modeling approaches have to allow for that. This area is of great interest because there are no approaches in literature so far.