In this paper Extreme Value Theory (EVT) and GARCH model are combined to estimate conditional quantile (VaR) and conditional expected shortfall (the expected size of a return exceeding VaR) so as to estimate risk of assets more accurately. This hybrid model provides a robust risk measure for the Tunisian Stock Market by combining two well known facts about security return time series: dynamic volatility resulting in the well-recognized phenomenon of volatility clustering, and non-normality giving rise to fat tails of the return distribution. We fit GARCH models to return data using pseudo maximum likelihood to estimate the current volatility and use a GPD-approximation proposed by EVT to model the tail of the innovation distribution of the GARCH model. This methodology was compared to the performances of other well-known modeling techniques. Results indicate that GARCH-EVT-based VaR approach appears more effective and realistic than Historical Simulation, static EVT, static normal and GARCH forecasts. The GARCH-EVT forecasts responded quickly to changing volatility, enhancing their practical applicability for Tunisian market risk forecasts.