On 11 April 2019, the then Governor-General Sir Peter Cosgrove called a Federal election for 18 May. According to the Newspoll published on 7 April, the two-party preferred vote reported Labor with an election-winning lead of 52% to 48% over the Coalition, with almost identical primary votes for the two major parties (39% each). Although there were some minor variations, all other polls, most pundits and the betting markets were all predicting an election win for Bill Shorten and Labor. That is not what transpired on election day. The eventual result was a return of the Scott Morrison Coalition government, with a twoparty preferred vote of 51.5% for the Coalition and 48.5% for Labor. Clearly, the polls were not able to accurately predict the election outcome. Despite supposedly being in an era of ‘big data’, prediction markets that make use of far larger sources of information also failed to predict the election outcome.

This paper summarises analysis of recently available longitudinal data to consider one potential aspect of the Australian election result variation between who respondents say they will vote for when asked in the lead-up to an election and who they end up voting for. Using linked ANUPoll data, we are able to track respondents at the individual level, with information on actual voting behaviour and voting intentions both at the time the election was called and a number of months before. In addition, we have information on some of the predictors of electoral volatility identified in the literature, as well as some predictors not used in previous analyses. In addition to highlighting a large degree of intraelection volatility that is partially explained by observable characteristics, the main conclusion from the analysis is that differences between polling results and the election outcomes were primary due to those who were undecided in the lead-up to the election, and those who said that they were going to vote for the non-major parties but swung to the Coalition.