TailRisk-Trans: A Transformer-Based Dynamic Tail-Risk Prediction Model with Extreme-Event–Aware Attention for Financial Markets

Abstract

Accurate modeling of tail risks such as market crashes, volatility spikes, and extreme downside events is crucial for portfolio management, quantitative risk control, and regulatory stress testing. Traditional econometric models, including GARCH and EVT-based hybrids, often struggle to capture the nonlinear dependencies and regime shifts inherent in modern high-frequency markets. To address these limitations, we propose TailRisk-Trans, a unified Transformer-based tail-risk prediction framework designed to dynamically forecast Value-at-Risk (VaR), Expected Shortfall (ES), and Conditional Value-at-Risk (CVaR). The model incorporates four key components: a financial data preprocessing layer that integrates microstructure features, derivative-implied risks, and macroeconomic factors; a Market Transformer Encoder capable of capturing long-range temporal dependencies; a Tail-Risk Prediction Head that jointly predicts multiple risk metrics; and an Extreme-Event-Aware Attention mechanism that adaptively increases sensitivity to volatility spikes and abnormal distributional shifts. Experimental evaluations on multi-market datasets-including equity indices, index futures, and volatility indices-demonstrate the superior performance of TailRisk-Trans under both normal and turbulent market regimes. Compared with the strongest baseline model, Transformer-TS, TailRisk-Trans reduces the 99 percent Value-at-Risk (VaR) violation rate from 4.12 percent to 3.47 percent , representing a 15.8 percent improvement in tail-event compliance. In addition, TailRisk-Trans attains the lowest quantile loss (2.684) and the most favorable Expected Shortfall score (3.892), confirming enhanced sensitivity to shifts in tail-risk distributions. These results highlight the practical applicability of TailRisk-Trans for quantitative trading strategies, real-time risk monitoring, and regulatory early-warning systems.