Value at Risk models with long memory features and their economic performance

We study alternative dynamics for Value at Risk (VaR) that incorporate a slow moving component and information on recent aggregate returns in established quantile (auto) regression models. These models are compared on their economic performance, and also on metrics of first-order importance such as violation ratios. By better economic performance, we mean that changes in the VaR forecasts should have a lower variance to reduce transaction costs and should lead to lower exceedance sizes without raising the average level of the VaR. We find that, in combination with a targeted estimation strategy, our proposed models lead to improved performance in both statistical and economic terms.