In this paper, Extreme value theory (EVT) is applied in estimating low quantiles of P/L distribution and the results are compared to common VaR methodologies. The fundamental theory behind EVT is built, and peaks-over-threshold method is used for modeling the tail of the distribution of losses with Generalized Pareto Distribution (GPD). The different VaR methods are then compared using backtesting procedures. Practical issues such as time varying volatility of returns, and multivariate time series (portfolio of financial instruments) are covered.