Data Mining and Machine Learning for Political Events: An Application to Produce Indicators for Presidential Systems

This paper presents the creation of a novel dataset on turnover and resignation calls in 12 Latin American cabinets from the mid-1970 to date. The information on resignation calls is entirely novel and constitutes a relevant empirical contribution to the study of presidential politics and public opinion. We focus on the creation of the dataset using optical character algorithms on archives and press reports together with machine learning models. The trained models consist of a set of supervised and semisupervised models for almost 50 years. Subsequently, we provide a number of validity and plausibility checks to cross-validate our data by comparing it with similar existing sets and offering practical applications on ministerial survival and salience issues during the period.