Convolutional neural networks to signal currency crises: From the Asian financial crisis to the Covid crisis

Currency crises are recurrent events in economic history. They were particularly frequent during the 1980s and 1990s, reflecting diverse underlying causes, and have continued to occur in the early decades of the 21st century. This paper proposes a unified model to examine recent crises across 60 countries between the Asian crisis and the Covid-19 pandemic, including the 2008 global financial crisis and the 2014–2016 commodity-related tensions. The objective is to develop a robust early warning system capable of identifying potential currency crises within a two-year horizon, regardless of their origins. We assess several state-of-the-art machine-learning architectures used in financial forecasting, going beyond conventional econometric benchmarks. For the first time in this literature, particular attention is given to convolutional neural networks, originally designed for image recognition, offering an innovative perspective for the analysis of macro-financial vulnerabilities. The results indicate that CNNs generate more accurate warning signals than other competitive models, such as long short-term memory networks, detecting 24 out of 27 crises in the sample. Moreover, the convolutional-based analysis replicates well-established empirical regularities, assigning varying importance to indicators across subperiods. While the collapses observed between 2014 and 2016 appear primarily driven by domestic macro-financial deterioration, the 2008 and Covid-19 crises are more closely linked to global or US factors.