Extreme Value Theory (EVT) is one of the most commonly used approaches in finance for measuring the downside risk of investment portfolios, especially during financial crises. In this paper, we propose a novel approach based on EVT called Uncertain EVT to improve its forecast accuracy and capture the statistical characteristics of risk beyond the EVT threshold. In our framework, the extreme risk threshold, which is commonly assumed a constant, is a dynamic random variable. More precisely, we model and calibrate the EVT threshold by a state-dependent hidden variable, called Break-Even Risk Threshold (BRT), as a function of both risk and ambiguity. We will show that when EVT approach is combined with the unobservable BRT process, the Uncertain EVT's predicted VaR can foresee the risk of large financial losses, outperforms the original EVT approach out-of-sample, and is competitive to well-known VaR models when back-tested for validity and predictability.