We report on an eight-year longitudinal study of predicting the outcome of elections based on party mentions in tweets. Five Dutch national elections for the parliament and senate between 2011 and 2019 were examined. Configurations with four parameters were tested. For three elections, reasonably accurate predictions can be obtained that are under twice the error of the classic polls, but only after post-hoc optimization. When the same optimal parameter configuration is used for all elections, the results worsen.
|Title of host publication||Social Informatics|
|Subtitle of host publication||SocInfo 2019|
|Place of Publication||Urbana-Champaign|
|Publication status||Published - 11 Nov 2019|
|Name||Lecture Notes in Computer Science|
Sanders, E., & van den Bosch, A. (2019). A Longitudinal Study on Twitter-Based Forecasting of Five Dutch National Elections. In Social Informatics: SocInfo 2019 (Vol. 11864, pp. 128-142). (Lecture Notes in Computer Science; Vol. 11864). Springer. https://doi.org/10.1007/978-3-030-34971-4_9