Purpose. We present an approach for forecasting mental health conditions and emotions of a given population during the COVID-19 pandemic in Argentina based on social mediacontents. Design. Mental health conditions and emotions are captured via markers, which link social media contents with lexicons. First, we build time seriesmodels that describethe evolution of markers, and their correlation with crisis events. Second, we use thetime seriesforforecastingmarkers, andidentifyinghigh prevalence points for the estimated markers. Findings. We evaluated differentforecasting strategies that yielded different performance and capabilities. In the best scenario, high prevalence periods of emotions and mental health issuescan be satisfactorily predicted with a neural network strategy, even atearly stages of a crisis (e.g., a training period of 7 days). Originality. Although there have been previous efforts to predict mental states of individuals, the analysis of mental health at the collective level has received scarce attention. We take a step forward by proposing a forecasting approach for analyzing the mental health of a given population at a larger scale. Practical implications. This work contributes to a better understanding of how psychological processes related to crises manifest in social media, and this isa valuable asset for the design, implementation and monitoring of health prevention and communication policies.