Attention Winged Deep Neural Early Warning System: How Does it Perform in Signaling Sovereign Crises for China?

Abstract

This study efforts to construct an effective early warning system (EWS) to predict sovereign crises for China and distinguish different levels of leading determinants, such as macro-economic fundamentals and risk transmission factors, impact significance in signaling the volatility quantified crises. We contribute the literature for predicting the sovereign crisis in three aspects of first defining ‘Chinese-style’ sovereign crisis by referring to the national bond index high volatile state regardless of the subjectivity from rating agency judging and the verbosity from threshold selecting criteria for each determinants, first technically quantifying contagious factors given specific crisis origins by multiplying the SWARCH model estimated high volatility probabilities of contagious sources to the DCC-GARCH model estimated correlation coefficients and first implementing the statistical hypothesis test on stylized attention based neural networks to judge the inferred contributing degree’s credibility. We find the regime classification regime metric is essential to classify crises and in the predictive models horse-race, the attention mechanism based bidirectional LSTM networks outperforms others. For China, the gold price, constant price for real estate in GDP and technically quantified contagious index for between oil and CDB index are top three leading indicators drawn by the attention at appropriate significant levels.

Keywords: Early warning system, Volatility defined sovereign crises, Technically quantified contagious factors, Attention based neural networks.

JEL Classification: C45,G01