Comparison of different question formats eliciting point predictions

Survey questions that elicit point predictions regarding uncertain events face an important challenge as human forecasters use various statistics to summarise their subjective expectations. In this paper, we take up the challenge and study whether alternative formulations of the questions used to elicit point predictions can be successful in driving forecasters towards reporting a particular central tendency (median or mean) of their subjective expectations distribution. We set up a laboratory experiment in which the participants act as forecasters and are asked to predict the next realisation of iid random draws coming from an objectively known distribution. We elicit the subjects' point predictions in four treatments, in which we ask for either (1) a "guess" of the next draw, as is standard in survey measures, (2) a "guess" as close as possible to the next 6 draws, and (3) the mean, or (4) the median of the next six draws. We then compare the predictions reported in the different treatments and their proximity to the three main central tendencies (mean, median, mode) of the objectively known distributions. We also investigate the cognitive process that affects the production of point predictions. We find that the majority of predictions in the two guess treatments, (1) and (2), are close to the mode. In treatment (2) ("one guess for the next six draws"), the forecasters report the mean and the median more often in comparison to (1) ("guess for the next draw"), but the mode remains the central tendency around which most of the predictions are located. In treatments (3) and (4), we find that forecasters adjust the point they report in the direction of a particular central tendency when specifically asked to report the mean or the median. Adjustments are more precise for forecasters with higher measures of numeracy and for those who have more experience. However, numeracy has no explanatory power when the forecasters are asked to report a "guess for the next draw" in treatment (1) which suggests that forecasters have different ways to summarise a distribution.