Early warning system for banking crises: an approach based on multinomial models.

Summary This thesis is an extension of the Early Warning Systems of banking crises based on a multinomial logit econometric approach. The objective of the thesis is twofold: in a first part, we analyze the concept, the determinants and the theoretical foundations of the banking crisis in a general context. In this way, we define a framework of analysis conducive to the implementation of a crisis prevention policy. In a second part, based on this analytical framework, we contribute to the prediction techniques of banking crises. The originality of the thesis lies in the use of multinomial logit models as a tool for predicting banking crises. From two empirical studies, we highlight three main results: first, we show that taking into account the specificities of banks in terms of accounting ratios improves the prediction of banking crises. Second, we detect the presence of a bias related to the identification of crises. This bias is revealed when we take into account the periods that precede or follow a crisis episode, which are neither quiet periods nor crisis periods. Third, we propose a framework for monitoring banking systems that consists of assigning scores in terms of the level of fragility. On this basis, we show, for example, that the most severe banking crises that occurred in Brazil and Mexico in 1994 and in Southeast Asia in 1997, follow periods of very high fragility.