In this article, we argue that electoral context affects the projection mechanisms inherent in polling. This insight applies both to the estimation of party vote shares by pollsters and to poll-driven substantive political expectations. To test this argument, we analyse 794 in-campaign polls covering the UK's 21 post-war general elections, as well as an updated version of Jennings and Wlezien's (2018) international polling dataset. We demonstrate that the election level houses a substantial portion of the observed variance in polling error. This finding is valid across several modelling approaches and a range of measures of polling accuracy both within and beyond the UK. Within the UK, we show that the election level is a particularly important locus of variance when it comes to analysing whether polls give rise to misleading substantive expectations about election results.

No abstract

Despite a wealth of literature existing with regards to electoral prediction, no pieces exist that address whether certain elections are more predictable than others. Indeed, the existing literature invariably focuses upon the accuracy and efficacy of the methods used to predict elections, without addressing the nature of the elections that are being predicted. Consequently, a fundamental avenue of enquiry remains unexplored within the current literature, resulting in a stark and surprising lacuna and rendering the dominant paradigm problematic. To remedy this deficiency, this paper addresses whether certain elections are more predictable than others and ascertains which characteristics cause their predictability to vary. Owing to both temporal and spatial limitations, electoral predictability is addressed with a specific focus on British national elections from 1945 to 2017. In order to achieve its aims, this paper engages in a series of bivariate analyses between measures of predictive accuracy and electoral characteristics. Specifically, the correlation between the total dichotomous and total mean absolute error exhibited by pre-election polls conducted for the twenty-three studied elections and six specified electoral characteristics is ascertained and utilised to draw conclusions. Ultimately, this paper finds that certain elections are indeed more predictable than others, with certain characteristics having a greater impact on the studied measures of predictability than others. However, due to both the small size of the utilised data set and its limited scope, it is accepted that the conclusions drawn are only indicative in nature. Despite this, owing to the unique nature of this paper, it is hoped that the conclusions drawn will provide a foundation for the future exploration of this hitherto unexplored avenue of enquiry to further our collective understanding of the predictability of elections.