This thesis aims to study the results of applying different models to estimate Brazilian stock volatilities. The models are applied to six series of daily returns, and each series has 1200 days. We studied first the series` main statistical features: Stationarity, unconditional distribution and independence. We concluded that the series are mean stationary, but there was no conclusion on variance stationarity, in this first analysis. Return distribution is not normal, because of the high kurtosis. Returns showed time dependence, linear and, mainly, not linear. We modeled the linear dependence, and then applied ten different volatility models, in order to try to capture the non linear dependence. We evaluated the different models, in sample and out of sample, by analyzing their residuals and their forecast errors. The results showed that the less sophisticated models tend to give a worst representation of the data generating process; they also showed that the less parsimonious models are difficult to estimate, and their results are not as good as we could expect from their sophistication. We used the ten models` volatility forecasts to estimate value-at-risk (VaR) and two methods to estimate the residual distribution quantiles: empirical distribution and extreme value theory. The results showed that the less sophisticated models give better VaR estimates. This is a consequence of the variance non stationarity, that became apparent along the thesis.