DLST: A Novel Framework for Market Risk Stress Testing via EXformer

Stress testing plays an important role in risk management. It is commonly used to evaluate the risk tolerance of financial institutions under extreme market conditions, providing early warning and responses to potential adverse scenarios, thereby enhancing their overall risk management capabilities. However, setting reasonable stress scenarios in the process of stress testing is often the critical challenge. To address this, we propose a novel market risk stress testing framework called DLST, which integrates our proposed deep learning model with stress testing econometric models to enhance the sensitivity of market risk stress testing to changes in financial market conditions. Specifically, we propose EXformer, which is more suitable for situations where a variable in the financial market is highly influenced by many exogenous variables. By revising embedding layers, the Transformer can obtain exogenous information to enhance the prediction of endogenous variables, thereby generating more realistic stress scenarios and increasing the reliability of stress testing models. Moreover, by integrating dilated causal convolution and ProbSparse self-attention, it not only improves efficiency and considers temporal causality but also captures long-term dependencies effectively. Experimental results demonstrate that EXformer outperforms existing state-of-the-art models on seven real-world datasets and exhibits notable generality and scalability.