Origins of Presidential poll aggregation: A perspective from 2004 to 2012

US political reporting has become extraordinarily rich in polling data. However, this increase in information availability has not been matched by an improvement in the accuracy of poll-based news stories, which usually examine a single survey at a time, rather than providing an aggregated, more accurate view. In 2004, I developed a meta-analysis that reduced the polling noise for the Presidential race by reducing all available state polls to a snapshot at a single time, known as the Electoral Vote estimator. Assuming that Presidential pollsters are accurate in the aggregate, the snapshot has an accuracy equivalent to less than ±0.5% in the national popular-vote margin. The estimator outperforms both the aggregator FiveThirtyEight and the betting market InTrade. Complex models, which adjust individual polls and employ pre-campaign “fundamental” variables, improve the accuracy in individual states but provide little or no advantage in overall performance, while at the same time reducing transparency. A polls-only snapshot can also identify shifts in the race, with a time resolution of a single day, thus assisting in the identification of discrete events that influence a race. Finally, starting at around Memorial Day, variations in the polling snapshot over time are sufficient to enable the production of a high-quality, random-drift-based prediction without a need for the fundamentals that are traditionally used by political science models. In summary, the use of polls by themselves can capture the detailed dynamics of Presidential races and make predictions. Taken together, these qualities make the meta-analysis a sensitive indicator of the ups and downs of a national campaign-in short, a precise electoral thermometer.