The most recent International Journal of Forecasting includes a paper titled “Forecasting election results by studying brand importance in online news.” by Andrea Fronzetti Colladon of Universita degli Studi di Perugia, Italy. In the paper the author analyzes the use of a Semantic Brand Score (SBS) for textual data to forecast elections based on online news. The SBS has three dimensions:
- Brand prevalence, the frequency of appearance of the brand in discourse,
- Brand diversity based on the broadness of other terms co-mentioned with the brand, and
- Brand connectivity, which measures the degree to which the brand name is linked to other terms that are not linked to each other.
Prevalence is calculated globally and seems similar to term frequency used in semantic analysis. Diversity and Connectivity are computed as graph metrics after transformation of the online news document corpus into a network of word co-occurrences.
SBS was used in a backtest for four elections in Italy, 3 in 2016 and 1 in 2018. SBS had it’s lowest error 1 week prior to each election and was more accurate when there were fewer alternatives in the election. One difference between the Italian system and the American system is a blackout of polling in the final 2 weeks leading up to elections. It would be interesting to see if the results for SBS as a predictor would transfer to a system where polling continues until the election.
The published results indicate that SBS is fairly consistent in terms of outcome rankings even if the percentages are off. It may be useful as a predictor in situations where there is a dearth of polling. My impression is that it’s not a strong predictor by itself but it would be a good complement to other prediction models. I’ll look forward to seeing if there are follow-on papers that extend the research as suggested by the author. I’d be especially interested to see the method applied to other elections.
One minor technical remark: early in the paper there’s a conjecture for two possible mechanisms for the relationship between SBS and popularity. One is the possibility that consumers of online news are primed to think more about political brands with higher SBSs. The other is the possibility that parties that are ahead in the polls have more attention and coverage which could lead to higher SBS. The author identifies this as something to investigate in future research. Perhaps a technique such as Structural Equation Modeling could be used to analyze the degree to which standing in the polls mediates the relationship between SBS and popularity.