The existence of fat tails is one of the striking stylized facts of financial returns distributions. This fact makes the use of traditional parametric models for Value at Risk (VaR) estimation unsuitable for the estimation of low probability events (1% or less). This is because traditional models are based on the conditional normality assumption for financial returns distributions, making them unsuitable to predict the actual probabilities of occurrence of atypical returns. The main purpose of this dissertation is to investigate the performance of VaR models based on Extreme Value Theory (EVT), and to compare them to some traditional models. Two classes of models are investigated. The first class is based in an unconditional model, which characterizes the long-term behavior of the time series of returns. The other class of models is a conditional one, which incorporates the short- term behavior of the return series, characterized by the strong dependency observed on the conditional variance of the returns.Both models were applied to four representative time series of the Brazilian stock market: The Bovespa Index, Future of Bovespa Index, Telesp stocks and Petrobrás stocks. The results indicates that EVT based models are suitable for low probability VaR estimation. Besides that, its possible to conclude that the conditional model is more appropriate for crisis periods, because of its capacity to quickly respond to volatility changes. Alternative risk measures are also used, to give estimates of the losses magnitudes in the case of VaR violation.