2014 Election forecast - a post-election analysis

Abstract

General elections are held every five years in South Africa. During the 12 to 24 hour period after the close of the voting booths, the expected final results are of huge interest to the electorate and politicians. In the past, the Council for Scientific and Industrial Research (CSIR) has developed an election forecasting model in order to provide the media and political analysts with forecasts of the final results during this period of peak interest. In formulating this model, which forecasts the election results as the results from voting districts (VDs) become available, some assumptions had to be made. In particular, assumptions were made about the clustering of previous voting patterns as well as the order in which VD results are released. This election forecasting model had been used successfully for a number of elections in the past and in these previous elections, with around 5%-10% of the results available, the predictions produced by the model were very close to the final outcome, particularly for the ANC, being the largest party. For the 2014 national election, however, the predictions, with close to 50% of the voting district results known (equivalent to an estimated 40% of the total votes), were still not accurate and varied by more than 1% for both the ANC and the EFF. This paper outlines a post-election analysis to determine the reasons for these discrepancies and how they relate directly to the model assumptions. The aim is to highlight how practical realities can affect the assumptions and consequently their impact on the forecasted results. Reference is made to previous election forecasts and the 2014 post-election analysis is presented.

From SA to the USA: Election forecasting

“…This election night model worked well in the South African context, except when certain model assumptions were violated, as in 2014 (see [27]). A summary of most of the results up to the 2014 elections, as well as issues regarding the model assumptions, can be found in Ittmann et al [27].…”

“…This election night model worked well in the South African context, except when certain model assumptions were violated, as in 2014 (see [27]). A summary of most of the results up to the 2014 elections, as well as issues regarding the model assumptions, can be found in Ittmann et al [27]. Typically the prediction results at a municipal level are less accurate than those at a national or provincial level since one is working with a smaller sample of voting districts, but the findings presented in Ittmann et al [27] were used to make improvements to the 2016 municipal election clusters with the result that better prediction accuracy was achieved.…”

“…Election night forecasting models generally become a source of insightful information during election coverage in the media while the election outcome is still unknown. Countries including the UK [8,9,37,38], USA [2,10], Spain [37], South Africa [19,20,27], New Zealand [36,37], Australia [37], Austria [25], Sweden [37] and Ireland [37] have, across a number of elections, made use of election night forecasting models to infer an election outcome during hours (and sometimes days) of uncertainty around the final outcome of an election. These models seek to provide an accurate forecast of an election outcome from a small but reasonable sample of the released results.…”

“…A summary of most of the results up to the 2014 elections, as well as issues regarding the model assumptions, can be found in Ittmann et al [27]. Typically the prediction results at a municipal level are less accurate than those at a national or provincial level since one is working with a smaller sample of voting districts, but the findings presented in Ittmann et al [27] were used to make improvements to the 2016 municipal election clusters with the result that better prediction accuracy was achieved. Overall, the 2016 municipal predictions for the metros' PR votes proved to be a great success with the model accurately predicting that coalition governments would take power in the Johannesburg, Tshwane (Pretoria), Ekurhuleni and Nelson Mandela Bay (Port Elizabeth) metros; that the DA would be the largest party in the latter metro; that the Democratic Alliance (DA) would consolidate their control over Cape Town; and that the African National Congress (ANC) would retain control over eThekwini (Durban), Buffalo City (East London) and Mangaung (Bloemfontein).…”

“…Since 1999 an election forecasting team from the Council for Scientific and Industrial Research (CSIR) have consistently, during every election, used an election night forecasting model to forecast the final election outcome using the first batch of voting district results as they are released by the Independent Electoral Commission (IEC). This model has performed well except when certain model assumptions were violated, as in 2014 (see [27]). After successfully predicting most of the metro (metropolitan region) results correctly in the 2016 municipal elections, the team felt that the model could possibly be adapted to work outside of SA and decided to test the model on the 2016 USA presidential elections.…”