This paper assembles our results from several lines of research on volatility forecasting for risk measurement applications. We examine volatility prediction from ARMA-GARCHclass models estimated using the quasi-maximum likelihood (QML) and robust bounded M method, and we illustrate the influence of the estimation method on value at risk (VaR) in the presence of outstanding observations. We apply the Monte Carlo to compare the results, assuming several fractions of outstanding observations. We explore the effect of outstanding observations on risk measurement for ARCH-GARCH-class models estimated with QML and robust BM estimators, and for conditional autoregressive value at risk (CAViaR), based on regression quantiles.