Human touch puts local pollsters way ahead
FROM a trendy office in the run-down Cape Town suburb of Salt River, a small South African software company is accurately predicting some of the biggest events in the world.
BrandsEye, with not much fuss, correctly called the Brexit result .
But the company made bigger waves by correctly predicting that Donald Trump would move into the White House. Major pollsters got it wrong, as did media houses.
BrandsEye even made the news. This week a CNN headline read: “Psychic software? How a small social media company predicted Donald Trump’s victory”.
The company, which primarily focuses on providing clients with real-time media analysis to manage their brands, is inadvertently reshaping how polling is done and how social media is used.
It has set itself apart by combining a crowdsourcing model, which uses humans to gauge conversational sentiment, with traditional artificial intelligence that collects relevant social media feeds.
“Our crowdsourcing approach of using humans to understand what people are saying and what the emotions are around those thoughts is really important,” said the company’s founder, Craig Raw.
“Most of our competitors go at things with just an algorithm.
“[But] algorithms don’t understand things like sarcasm, humour, mixed meaning and all of the nuances that come in, particularly when you are trying to squeeze your message into 140 characters.”
Regular pollsters collect information from around 1 000 to 2 000 people on average for their polls.
During the US election BrandsEye collected 37 million social media conversations from four million users.
Thousands of people throughout the world were crowdsourced to do human analysis on 200 000 individual messages.
BrandsEye CEO JP Klopper said traditional polls were flawed in that they contained both question and sample bias, but social media conversations were often unsolicited and therefore had fewer filters.
A senior data scientist at market research firm TNS, Kyle Findlay, said the model used by BrandsEye was simple and effective, but that even more could still be done in the field to improve accuracy.
“Accuracy levels for most algorithms is around 60% to 70%. So generally speaking, algorithms are not very good. To get them to be accurate, you have to have humans involved and that is the difference with them,” said Findlay.
Social media expert Toby Shapshak said he believed the results showed that social media was playing an interesting and evolving role in people’s lives, and that it now had a cachet it did not before.
Last year BrandsEye looked at the #FeesMustFall protests and found that the people were not divided on race but rather on the issue.
In a separate survey, it found that Capitec had the most positive sentiment around it and that positive perception of the big four banks was overblown.
The company plans to look at the upcoming French elections next.