Azra Hoosen | ah@radioislam.co.za
30 May 2024 | 11:00 CAT
2 min read

The Council for Scientific and Industrial Research (CSIR), in collaboration with the South African Broadcasting Corporation, is set to once again employ its election night prediction model for the upcoming 2024 national and provincial elections. This advanced model, first introduced during the 1999 general elections, has been a staple in predicting election outcomes for the past 25 years.

The CSIR has utilised a statistical approach to forecast results at various levels. The model’s foundation lies in analysing voter behaviour patterns and the sequence of vote result announcements, leveraging statistical clustering methods to group voters or voting districts based on historical data.

CSIR’s Mahlatse Mbooi, senior data scientist, said the predictive tool has been quite accurate, with an error margin of about 2%. “Only a few areas were not very accurate, but generally, it has performed quite well over the years. No changes have been made to the model; only the input data has changed,” she said.

According to Mbooi, There has been a growing interest in analysing social media discussions surrounding elections. In previous years, they incorporated social media data into the predictive process. However, she emphasised it’s important to note that conversations on social media do not accurately reflect voter behaviour at polling stations. The sentiments and opinions expressed online often differ significantly from actual voting patterns.

She pointed out that with regard to independent candidates, the model treats them as if they are political parties.

“What applies to the independent candidates is how the accuracy performs when it comes to small party predictions, but since there is no previous data it will be hard to predict how they will perform in this elections,” she added.

According to CSIR Chief Executive Officer, Dr. Thulani Dlamini the CSIR’s election prediction model is not a polling system but a tool that uses statistical and mathematical analysis to forecast election outcomes. “It demonstrates how statistical clustering and certain mathematical algorithms can produce accurate predictions from a small sample of results. The model functions by reducing bias caused by the ‘non-randomness’ of incoming results, which stems from the sequence in which the results are received,” she explained.

LISTEN to the full interview with Ml Junaid Kharsany and CSIR’s Mahlatse Mbooi, here.