by , ,
Recent forecasting research has shown that a forecaster's past accuracy is predictive of her future accuracy. In settings, where participants report forecasts on several events and where some event outcomes materialize earlier than others, a forecaster's accuracy on early-closing events can be used to improve weighted forecast aggregation. In this paper, we propose and implement a new class of aggregation algorithms that uses a forecaster's accuracy without relying on any early-closing events. The key idea is to impute a forecaster's accuracy on open events using the average quadratic distance between her forecast and a "ground truth proxy." An example of such a proxy is the median of forecasts. We explain the theoretical motivation behind this approach and emphasize its practical effectiveness using data from a large-scale geopolitical forecasting tournament.
Forecast Aggregation using Imputed Accuracy J. Witkowski, P. Atanasov, L. H. UngarIn ICML'15 Workshop on Crowdsourcing and Machine Learning (CrowdML'15), 2015
Bibtex Entry:
author =	 {Witkowski, Jens and Atanasov, Pavel and Ungar, Lyle H.},
  title =	 {Forecast Aggregation using Imputed Accuracy},
  booktitle =	 {ICML'15 Workshop on Crowdsourcing and Machine
                  Learning (CrowdML'15)},
  year =	 {2015},
  month = 	 {July},