A framework for improving electoral forecasting based on time-aware polling

Abstract

The prediction of opinion distribution in real-world scenarios represents a major scientific challenge for current social networks analysis. In terms of electoral forecasting, we find several prediction solutions that try to combine statistics with economic indices, and machine learning, like multilevel regression and post-stratification (MRP). Nevertheless, recent studies pinpoint toward the importance of temporal characteristics in the diffusion of opinion. As such, we take inspiration from micro-scale temporal epidemic models and develop an original time-aware (TA) forecasting methodology which is able to improve the prediction of opinion distribution in an electoral context. After a parametric analysis, we validate our TA method with pre-election survey data from three presidential elections (2012–2019) and the UK Brexit (2016). Benchmarking our TA method against two classic statistical approaches, like survey averaging (SA), and cumulative vote counting (CC), and the best pollster predictions, we find that our method is substantially closer to the real election outcomes. On average, we measure prediction errors of 9.8% (SA), 9.6% (CC), 5.1% (MRP), and only 3.0% for TA; these differences translate into increases of prediction accuracy of $$\approx 71\%$$ for the TA method (40% better than best pollsters). Moreover, TA does not require socio-economical contextual information, while the more complex MRP method depends on them for prediction. This work builds upon existing studies on the microscopic temporal dynamics of social networks and offers new insight on how macroscopic prediction can be improved using time-awareness.