Mortality models most often are used to make projections of life expectancy. A good mortality model should satisfy some desirability criteria (Cairns et al (2008)). Models should be robust which means that parameter uncertainty should be low and small changes in the data should not result in significant changes in the estimates of the parameters and in their interpretation. Most of the existing mortality models are not robust against outliers due to wars, pandemics etc. or so called "longevity outliers". This paper is not the first attempt to deal with outliers in mortality data. Hyndman and Ullah (2007) used a combination of robust, nonparametric statistics and functional data analysis in developing a method for projection of age-specific mortality rates observed over time. While their objective was to identify and remove outliers, we highlight the necessity of incorporating them into projections in order to capture, in a more realistically way, perturbations that may occur in the future. The main contribution of this paper is to utilize a highly robust estimator to minimize the effect of outliers on point forecasts of li ... mehrfe expectancy. We compare the point forecast accuracy and bias of seven stochastic models for life expectancy projection in the presence of outliers. Based on one-step ahead forecast errors we conclude that the Hyndman and Ullah method (2007) is the most accurate and the least biased and in the Lee-Carter family of models the Lee-Carter (1992) and the Cairns-Blake-Dowd (2006) produce the most accurate point forecast of life expectancy when death rates across outlying years are replaced by highly robust estimates.