Wisdom of the crowd: How to get rid of errors and biases at the aggregate level

In many experiments, it has been shown that the wisdom of the crowd can outperform individuals and sometimes even experts. In this thesis, I test two approaches to further improve the performance of four different mathematical aggregation models in forecasting football matches. Two unweighted models, the mean and median, and two weighted models, the Brier Weighted Model (BWM) and the Contribution Weighted Model (CWM). The BWM weighs forecasts of subjects based on individual expertise, while the CWM determines the weighting based on the contribution that a subject makes to the crowd’s expertise. The first approach improves the models by modifying the BWM and CWM. This is done by decreasing the power of past performance for the first 10 events, which reduces the problem of requiring a track record without decreasing the prediction accuracy. The second approach increases the forecasting accuracy of the mean, median and BWM significantly by providing subjects with additional information on what three other subjects estimated. I show that this increase in accuracy is due to subjects having more doubts, which decreases both the bias of estimating too extreme values as well as fans estimating too optimistically for their team. This decrease in biases overweighs the anchoring bias due to the displayed estimates. In addition, a more thorough analysis of the CWM shows that different than stated in Budescu & Chen (2014) the model seems to only partly identify experts based on individual expertise.