In this article, we study the point predictions that forecasters report when they are asked to predict the realisation of an iid random variable. We set up a laboratory experiment where the participants act as forecasters predicting the next realisation of random draws coming from different "objectively known" distributions which vary in the location of their central tendencies. As is standard in survey measures, the subjects in our experiment must report their best guess of the next draw as a forecast. We find that most of the forecasters report point predictions that are close to one of the three main central tendencies (mean, median or mode) of the distributions provided, with a majority corresponding to the mode. Our analysis also shows that when selecting a point prediction, people have in mind a numerical value (e.g. the mean or the mode) rather than a specific percentile of the underlying distribution. Only 5% of the forecasts reported during the experiment are based on a percentile while almost 60% are based on a numerical value.
subjective expectations forecasting eliciting point predictions experiment