I have attempted to develop a VaR model which assumes (Normal distribution). I have monthly hedge fund index returns and i have used a standard approach using a rolling window of 36 months to attain the desired parameters (total number of observations for each index 176). I have then adopted the back testing approach -quadratic probability score (QPS) method as a comparison of alternative models as per (Lopez, (1998)). Furthermore, i have also run thus far a GARCH (1,1) using MLE to model volitility and thus incorporated the GARCH into the VaR model. I thought that by using the GARCH output it should lead to better model performance, however, this has not been the case. Far from it!!Can anybody think of any reasons (apart from an error in calculation) i.e. is the data set too small? I can send the models over in excel if someone is willing to take a look. Cheers,Pierre