Diving into recession: the collective knowledge of online users as an early warning system for recessionary expectations

Abstract

As concerns about economic downturns manifest in online discussions, we investigate whether sentiment extracted from social media can serve as an early warning signal for recessionary pressures. Using a dataset of Twitter (X) posts related to economic prospects, we apply a range of sentiment analysis techniques, including a lexicon and rule-based method (VADER) and deep learning approaches (GPT and BERT). We assess the relationship between online sentiment and key recession indicators, such as the yield curve and GDPNow forecasts, using a combination of econometric and machine learning methods. In addition, we perform a comparative evaluation of sentiment classification techniques, incorporating both traditional models and deep learning architectures. Our results confirm that Twitter discussions precede changes in recessionary indicators and can thus provide forward-looking insights into economic sentiment. Furthermore, the comparative analysis reveals variations in sentiment detection across different methodologies, emphasizing the importance of selecting appropriate approaches in economic forecasting.