Data-driven decision-making is a new paradigm for solving various problems in a broad spectrum of domains. Politics and predicting the outcome of political elections is no exception, and interest is growing for both, scientific and professional communities in applying and exploiting the advantages of data analytics and machine learning algorithms. This paper strives to make several scientific and professional contributions in the domains. Traditionally, the most common approach to measure public opinion are sampling surveys that ask carefully defined questions to precisely selected samples of the population. However, such an approach has a high price: large investments of time, effort and money for researchers who design the survey, who collect the data, and the respondents who answers. The problem with surveys is the honesty of the respondents, as well as the sample. Furthermore, a number of elections outcomes forecasts recently differed from the final results, which shaken confidence in these kinds of forecasting. An alternative to such an approach has recently emerged with the potential to complement it or even to completely replace previously used research methods that would reduce costs for researchers and removed the effort for respondents by using social media data. In the domain of political campaigns, this potential is extremely large considering the fact that practically all political candidates and political parties use social networks in their campaigns. The research conducted here applies machine learning algorithms on data from social network Facebook user’s activity to develop predictive models of election outcomes. Main goal of the research is to compare their accuracy and reliability with models obtained by traditional public opinion polls. The research is conducted on French local elections. Four different approaches to machine learning are used to develop predictive models: an approach based on errors (artificial neural network algorithm), information-based approach (decision tree algorithm), a similarity-based approach (k-nearest neighbours) and an approach based on probabilities (Naive Bayes classifier). The evaluation and comparison of the models will determine the best machine learning algorithm for predicting election outcomes based on social data network Facebook. Obtaining equally effective predictive models while being faster and simpler access to data provides a significant scientific and professional contribution to research in this area.