Convolutional Neural Networks to signal currency crises: from the Asian financial crisis to the Covid crisis

Currency crises are a recurrent event in economic history. They were particularly numerous in the 1980s and 1990s, revealing a diversity of causes, and they have continued to occur in the first decades of the 21st century. In this paper, we will resort to a single model to simultaneously examine the recent crises in 60 countries between the Asian crisis and the Covid crisis, including the 2008 Global Financial crisis and the financial tensions of 2014-2016. The ultimate objective is to develop a robust early warning system capable of pointing out potential currency crises, regardless of their origins, within a two-year window before they occur. To achieve this, in addition to the literature benchmark, some state-of-the-art models used for forecasting financial asset prices and financial crises have been examined. For the first time in the literature, particular atten- tion has been paid to convolutional neural networks designed for image analysis, offering an innovative perspective for the analysis of currency crises. Our results show that these networks generate better warning signals than other robust candidates (long short-term memory neural networks for example) with 24 crises detected out of 27. The analysis of the results of the convolutional networks confirms empirical regularities identified in the literature, assigning significant weight to different indicators depending on the period under review. While the collapses observed between 2014 and 2016 are more likely to be the result of the deterioration of domestic macroeconomic and financial factors, the 2008 and Covid crises are rather attributed to global or US-related factors.