The Council for Scientific and Industrial Research’s (CSIR’s) election prediction model came within 1.45 percentage points of Wednesday’s election results in Ekurhuleni and Tshwane but underestimated the DA by 5.95 points in Johannesburg early this morning with more than two-thirds of voting districts declared.
The model also underestimated the DA in Cape Town, by 4.75 points, according to ITWeb’s artificial intelligence-assisted comparison of its municipal projections with Independent Electoral Commission (IEC) proportional representation (PR) vote data.
The comparison uses PR vote shares from the IEC extracts, rather than ward results or council-seat allocations. The IEC figures were captured close to the time of the CSIR snapshots, although the proportion of voting districts declared differed between metros.
The figures are an interim assessment of the model.
Developed in collaboration with the South African Broadcasting Corporation, the CSIR dashboard updates its forecasts as results are declared. The CSIR introduced the model during the 1999 general elections.
Poll position
Ahead of Wednesday’s local government elections, research company Ipsos said its survey found no single party commanded majority support nationally and it saw a significant decline in ANC support since 2021.
Ipsos says the DA appeared to have consolidated support beyond its 2021 and 2024 levels, with the party campaigning over an extended period, particularly in Johannesburg.
The research company adds that the MK Party, contesting a local government election for the first time, could replicate its 2024 national performance, while ActionSA’s increased support was driven primarily by Gauteng.
Hits and misses
This morning’s AI-assisted data extraction showed that the CSIR’s projection for the DA in Johannesburg, 34.5% against 40.45% in the IEC data, was 5.95 percentage points lower. It also overestimated ANC support, projecting 20.1% against the IEC’s 17.22%.
The CSIR’s ActionSA projection was considerably closer, at 17.1% compared with 16.53% in the IEC figures. The CSIR dashboard showed 45.8% of voting districts declared when this snapshot was taken at 06:29 on 6 November.
The gap was also evident in Cape Town, where the CSIR projected 56.7% for the DA, compared with 61.45% in the IEC data. It projected ANC support at 14.4%, against 9.82% in the IEC figures.
The model’s snapshot, taken at 05:59, showed 50.8% of voting districts declared. The IEC data extract used for the comparison was generated at 06:11.
Nearly there
Ekurhuleni was a much closer match. The CSIR projected 29.8% for the DA, compared with 29.44% in the IEC data.
The CSIR’s ANC projection was 22.4%, against 22.17%, while ActionSA was projected at 19.5%, compared with 19.92%. The model’s dashboard showed 59.7% of voting districts declared at 05:59.
In Tshwane, with 64.3% of voting districts declared, the CSIR projected 33.9% for the DA, against 32.45% in the IEC data. It slightly underestimated ActionSA, projecting 23.6% against 24.13%, and the ANC, at 20.5% against 21.55%.
The CSIR dashboard did not have a prediction for eThekwini at the time of the snapshot. The IEC extract, generated at 06:26, showed the MK Party leading the PR vote with 44.09%, followed by the DA at 20.46% and the ANC at 16.78%.
Under the bonnet
Yet, the CSIR says its model has consistently proved highly accurate, often producing stable projections once only a small percentage of districts has been declared and converging with final outcomes as more results come in.
The model groups voting districts with similar characteristics through statistical clustering techniques, “allowing it to generate projections before all voting districts have reported their results”, says the CSIR.
“The CSIR’s election prediction model demonstrates how scientific expertise can be applied to matters of national importance. By combining historical electoral data with advanced statistical modelling, we are able to provide the public with timely, scientifically informed projections as election results become available,” says Dr Jabu Mtsweni, executive cluster manager for NextGen Enterprises and Institutions.
The CSIR says the tool is “one of the few election prediction systems of its kind in Africa,” that is based on two principles: an analysis of historical voting behaviour patterns and an understanding of how voting district results are received and reported during counting.
Mtsweni adds that “importantly, these projections are not opinion polls or official election results, but independent statistical estimates based on the results declared at a particular point in time”.
The IEC is the only authority mandated to declare official election results.
Share
