Title: Model selection using predictive distributions of stochastic processes
Authors: Jonas Wallin - Lund University (Sweden) [presenting]
Abstract: Often when doing forecasting on real data, there is a range of possible models to use. Typically the ranking of the models is determined by the forecasting ability of the models, which in term determined by some scoring function. If this scoring function will, in the long run, select the true model, if available, is known as a proper scoring rule. This scoring function has a long history in the forecasting literature, especially in weather forecasting. If the distribution for one's observations has varying scaling, we show that it is important to take into consideration the result of the scoring rule. This is a characteristic that is overlooked in the scoring rule literature. Further, many of the popular scoring rules, like mse, mae and CRPS do not take scaling into account. We develop a set of new scoring rules and show that these rules take into account the scaling of the predictive distributions.