This paper compares the Value at Risk (VaR) forecasting performance of different quantile regression models to conventional GARCH specifications on the Nord Pool system price. The sample covers hourly data from 2005-2011. In order to identify significant explanatory variables, we use a linear quantile regression to characterize the effects of fundamental factors on the system price formations. From our analysis we are able to show how the sensitivity of the variables change over the range of price quantiles and detect how these sensitivities vary over the hours of the day. Our findings suggest that the demand forecast and the price volatility is the most important determinants of the price in the tails of the distribution. We use these variables in the further analysis and test the out-of-sample VaR performance of linear quantile regression, exponentially weighted quantile regression (EWQR) and conditional autoregressive value at risk (CAViaR) models on the system price. We extend the CAViaR models to account for asymmetrical response to returns and are innovative in including explanatory variables in the CAViaR specification. Our results show that the I-GARCHX CAViaR model with demand forecast as explanatory variable outperform the other models, and that CAViaR models in general perform well. The linear quantile regression with price volatility as explanatory variable also provides good results. The computational complexity of CAViaR models favors a linear quantile regression, so market participants have to make a tradeoff between the level of accuracy in the forecasts and the complexity of the model. Our findings are useful for producers, consumers and traders, as well as clearinghouses, as they provide an accurate measure of the price risk.