Predicting public opinion dynamics with the SCARDO-model

Over the past 20 years, the theory of agent-based social influence models has been actively developing, a trend which is associated with the need to explain opinion formation processes in the context of the digitalization of communication channels and the intensification of information exchange processes. However, the practical side of this theory remains poorly studied. The key reason for this is the difficulties in calibrating model parameters and thus constructing an empirical foundation. This paper validates the SCARDO-model of opinion formation using empirical longitudinal data from the social network VKontakte. The data include three opinion snapshots of a large-scale sample of VKontakte users and a snapshot of their friendship connections. The model parameters are calibrated on the first two snapshots, whereas the third one is used to check the accuracy of the model’s forecast regarding the populations of opinion fractions at the next time moment. The constant trend model serves as a benchmark. The analysis performed shows that, depending on the method of parameter calibration, the prediction of the SCARDO-model can be more or less accurate than those of the constant trend model. At the same time, changes in public opinion in the dataset at hand (despite being sufficient to calibrate the model parameters) are small from the macro-scale point of view and, as a result, the typical value of the forecast error does not exceed one percent of «votes».