In regression analysis, prediction intervals give a range of values that is likely to contain a new observation given the outcome of some predictor variables. A straightforward approach to obtain estimates of such intervals is to predict conditional quantiles of the response variable. In recent years, vine copula-based quantile regression has proven to be a flexible and competitive method in the field of quantile regression. In this thesis, we revisit D-vine copula-based quantile regression following Kraus and Czado (2017) to obtain prediction intervals for next-day air temperatures in Seoul, South Korea, given a set of predictor variables. In particular, we investigate the predictive ability of this method compared to linear quantile regression as a reference method. Here, performance is quantified as a scoring that rewards narrow intervals and adds a penalty if an observation misses the interval. At first, we compare three different approaches to select the training set for next-day predictions. In addition to an intuitive rolling window of successive days prior to the next day, we use a refined rolling window including days from previous years, and also a random selection of days. Overall, the refined rolling window showed best prediction results. Secondly, we investigate an extension of the training horizon for the refined rolling window and a threshold for the number of predictors included in the model. Especially for a larger training horizon, the D-vine quantile regression slightly outperforms the linear quantile regression.
«
In regression analysis, prediction intervals give a range of values that is likely to contain a new observation given the outcome of some predictor variables. A straightforward approach to obtain estimates of such intervals is to predict conditional quantiles of the response variable. In recent years, vine copula-based quantile regression has proven to be a flexible and competitive method in the field of quantile regression. In this thesis, we revisit D-vine copula-based quantile regression follow...
»