Research on systemic risk early warning of China’s banking system based on recurrent neural network: theoretical framework, measurement methods and empirical analysis

Abstract

: Early identification and precise early warning of systemic financial risks constitute the core challenge in maintaining financial stability. This study overcomes the methodological limitations of traditional linear econometric models by developing a nonlinear dynamic early warning framework based on recurrent neural networks (RNNs) to enhance risk identification accuracy and timeliness. Theoretical analysis demonstrates the inherent advantages of RNN models in capturing long memory characteristics, path dependence, and complex nonlinear correlations in financial time series. Using daily return data from major Chinese listed banks and securities firms from January 2007 to July 2022, we establish a dual–dimensional systemic risk measurement system incorporating conditional expected value at risk (ΔCoVaR) and marginal expected loss (MES). Improved LSTM and GRU architectures were designed to address gradient vanishing issues, with empirical results showing RNN models significantly outperforming traditional models in risk prediction. Key findings include: First, LSTM models achieve approximately 23% higher prediction accuracy for ΔCoVaR compared to traditional VAR models, identifying risk accumulation signals with 3–6 months of lead time on average; Second, bank size, leverage ratios, and maturity mismatch serve as core drivers of systemic risk, exhibiting significant nonlinear threshold effects; Third, RNN models demonstrated robust early warning capabilities during critical events such as the 2008 global financial crisis and the 2020 pandemic shock. These conclusions provide novel technical pathways for macroprudential regulation and offer crucial policy guidance for refining systemic risk monitoring frameworks.