The idea of utilizing data from social media for making predictions has generated great interest (Kalampokis et al., 2013; Schoen et al., 2013). The question is what can one predict based on such data? Prominent examples for such studies are prediction models that investigated the emotional constitution of people (Fernandez et al., 2012; Kross et al., 2013; Ortigosa et al., 2014), personal traits and characters (Kosinski et al., 2013), stock market behavior (Bollen et al., 2011; Siganos et al., 2014), election results (Alonso and Vilares, 2016; Tumasjan et al., 2011).

Further examples are consumer behavior (Ringelhan et al., 2015), public health (Sinnenberg et al., 2017), opinion flow (Wu et al., 2014), sharing cascades (Kupavskii et al., 2012; Cheng et al., 2014), account classification (Chu et al., 2010, 2012; Dickerson et al., 2014), conflicts among friends (Liu and Weber, 2014), demographics of users (Culotta et al., 2015), mental health (Guntuku et al., 2017), heart disease (Eichstaedt et al., 2015), tourism (information search and decision-making behaviors) (Zeng and Gerritsen, 2014), word-of-mouth (WOM) or consumer reviews (Zhang et al., 2012), box-office revenue of movies (Asur and Huberman, 2010), levels of rainfall (Lampos and Cristianini, 2012), earthquakes (Sakaki et al., 2010), theoretical implications introduced by social media (Kane et al., 2014). In Table 1 we provide a comprehensive overview of many important questions that have been studied using social media data. We would like to note that here we emphasized the “What to predict” aspect of these studies by highlighting the questions that have been addressed.

As one can see from Table 1 there are many different questions studied so far. In order to organize these publications, we introduce a taxonomy to categorize these publications according to a few major variables. In Figure 1 we give a graphical summary of our taxonomy. Overall, these questions fall into seven different fields (E, Economy; G, Geophysics; H, Health; M, Management; S, Sociology; Ps, Psychology; Po, Politology) covering almost all science areas. In this figure, we provide furthermore information about four additional layers, namely (I) the time horizon of the prediction (horizon) for making predictions about the future (F) or the present (P), (II) the level of prediction (level) for macro (Ma) and micro (Mi) level predictions, (III) the time of prediction (time) for batch (Ba) and real-time (Rt) predictions, and for (IV) making spatial (Sp) or non-spatial (Ns) predictions. Each of these layers will be discussed in the following sections.

One area missing from the above (see Figure 2) were studies in humanities. By performing a WoS search looking for articles containing the words Twitter/Facebook, humanities, and prediction/forecast we found no results. However, we found articles (54) searching for Twitter/Facebook and humanities. Interestingly, these articles are descriptive rather than predictive in nature. Examples for such studies are (Vainio and Holmberg, 2017). In Lee et al. (2017) and Vainio and Holmberg (2017) the authors studied who tweeted scientific articles with at least one Finnish author/co-author and that had high altmetric counts on Twitter and in Lee et al. (2017) the use of Twitter by scholars in the digital humanities was studied for informal scholarly communication. Those and similar papers performed a descriptive statistical analysis but no predictions were made.

There are two different types of prediction models used in the literature with respect to the prediction itself. The first type predicts the future and the second prediction type predicts the present. The former type is naturally understood because this is what is usually implied by a prediction or a forecast, namely that it should tell us something about the near or far future. For this reason, almost all of the above studies are from this type. However, the second type is unconventional because neither in classical statistics nor machine learning such predictions are made. An example in our context is the prediction of rainfall levels Lampos and Cristianini (2012). Here the idea is to use Twitter users as sort of social sensors that report real-world events instantaneously. Another example is the prediction of earthquakes (Sakaki et al., 2010). In the literature such predictions are called nowcasting or predicting the present (Schoen et al., 2013).

Another distinction in the predictions is with respect to the level of the prediction. The majority of articles makes predictions on a macro-level for which individual Twitter or Facebook users are irrelevant. Instead, what is important is the aggregation of users into categories. Examples for this is, e.g., predicting outcome of elections or box-office success of movies (Asur and Huberman, 2010; Alonso and Vilares, 2016; Tumasjan et al., 2011). In contrast, predictions on the micro-level make predictions for Twitter or Facebook users themselves. Examples are predicting the personality (Golbeck et al., 2011; Quercia et al., 2011; Hughes et al., 2012; Youyou et al., 2015) or human mobility (Jurdak et al., 2015).

The difference between batch and real-time models is that in the former case data are gathered off-line and then one prediction is made. In the latter case this process is iterated multiple times and data are generated on-line. Examples for batch predictions are election forecasts whereas real-time predictions forecast the political opinion continuously (Alonso and Vilares, 2016; Tumasjan et al., 2011). In general, the need for developing a real-time model depends on the application one is aiming at. For instance, if one intends to predict the outbreak of an epidemic of an infection disease this needs to be done in a real-time manner because there is not one scheduled event to occur one wants to predict but there is all the time a possibility for the outbreak to happen (Robillard et al., 2013; Santillana et al., 2015). Another example is the prediction of stock market values (Bollen et al., 2011; Siganos et al., 2014).

A final distinction of prediction models relates to non-spatial vs spatial predictions. A non-spatial prediction makes a forecast for the population as a whole, e.g., the outcome of an election (Alonso and Vilares, 2016; Tumasjan et al., 2011). In contrast, a spatial prediction makes a forecast for, e.g., all municipalities of a country. In this sense predictions in the former case can be considered as scalar whereas in the latter case they are multivariate. In order to accomplish a spatial prediction, usually information about the geolocation of the users is utilized. This information may be either directly available, or needs to be inferred from the content of the microblogs.

When collecting the articles for this review we noticed that despite the fact that all considered publications utilize prediction models, only a small fraction of these make an attempt to ensure the statistical soundness of the models. As a simple indicator for this omission we searched the WoS for articles containing the keywords Twitter or Facebook and for articles that contain the keywords Twitter and cross validation or Twitter and resampling (similarly for Facebook). The result of these searches is shown as a scatter plot in Figure 1D. The shown pairs correspond to the same publication year and y-axis label SM & CV is an abrieviation for our second search query. This figure confirms our perception indicating that only a small fraction of all articles applies resampling methods in order to quantify the uncertainty in the data and to guard against overfitting. Given the fact that the analyzed social media data are “big,” resampling methods can always be applied. Overall, this indicates a possible problem that would require further analysis.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.