Dynamic selection and combination of conditional quantile forecasts, with application to value-at-risk modeling

We introduce an effective and computationally fast approach to combine conditional quantile forecasts. The approach uses the information of the relevant loss function for the quantile problem associated to each candidate model in order to define forecast combination weights in a dynamic fashion. Two important advantages of the proposed method are that i) does not require numerical optimization of the combination weights, which facilitates implementation when a large cross section of individual forecasts is considered and ii) the aggressiveness in the allocation across alternative forecasts and the trimming of worse forecasts can be easily calibrated with a single parameter. An empirical implementation of the method based on a large dimensional data set with 50 assets and on value-at-risk (VaR) forecasts obtained with a set of 16 alternative candidate models shows that the portfolio VaR forecasts based on the proposed method are accurate and outperform that of individual models in many instances.