Dynamic Monitoring of Banking Systemic Risk: An Analysis Based on Smooth Quantile-LSTM Model

Abstract

Preventing systemic risk is a core task of macroprudential supervision, and accurately measuring the marginal contribution of individual banks to systemic risk remains a key challenge. To address the estimation bias of traditional parametric models in characterizing extreme tail risk, this paper constructs a Smooth Quantile Long Short-Term Memory (SQ-LSTM) model and conducts an empirical analysis using daily return data of 14 European Global Systemically Important Banks identified by the Financial Stability Board from 2008 to 2021. The model introduces the Horowitz smooth kernel to reformulate the traditional quantile loss function so that it is differentiable over the entire domain, employs LSTM gating mechanisms to capture long-term dependencies and regime-switching features of financial time series, and uses a multi-output neural network to simultaneously estimate VaR and CoVaR for all institutions. The empirical results show that SQ-LSTM is highly responsive to extreme events, with ΔCoVaR peaks closely coinciding with major shocks such as the 2008 financial crisis, the 2011 European debt crisis, Brexit in 2016, and the COVID-19 pandemic in 2020; risk rankings remain highly stable throughout the sample period, with Spearman correlation coefficients consistently above 0.80; and out-of-sample rolling forecasts with Diebold–Mariano tests confirm that, at the 1% extreme quantile level, SQ-LSTM achieves significantly higher prediction accuracy than the benchmark Q-LSTM model. This study provides a distribution-free, robust, and sensitive quantitative framework for identifying systemically important institutions and dynamically monitoring systemic risk.