Multi-cycle forecasting of congressional elections with social media

Abstract

Twitter has become a controversial medium for election forecasting. We provide further evidence that simplistic forecasting methods do not perform well on forward-looking forecasts. We introduce a new estimator that models the language of campaign-relevant Twitter messages. We show that this algorithm out-performs incumbency in out-of-sample tests for the 2010 election on which it was trained. That success, however, collapses when the same algorithm is used to forecast the 2012 election. We further demonstrate that volume-based and sentiment-based alternatives also fail to forecast future elections, despite promising performance in back-casting tests. We suggest that whatever information these simplistic forecasts capture above and beyond incumbency, that information is highly ephemeral and thus a weak performer for future election forecasts.