B0585
Title: Sequential model confidence sets
Authors: Sebastian Arnold - University of Bern (Switzerland) [presenting]
Johanna Ziegel - University of Bern (Switzerland)
Abstract: In most forecasting situations, a whole set of different and possibly competing models are given. Naturally, selecting the best models amongst all available ones is desired, where the best is understood in terms of appropriate loss functions. The model confidence set (MCS) algorithm by a prior study, provides a powerful solution to this problem. However, the MCS algorithm only allows for inference over an evaluation period that is fixed in advance. The MCS algorithm is adapted and extended since forecasting and forecast evaluation are inherently sequential tasks: data is collected and accumulated sequentially over time and want to draw inferences on a regular basis, as, e.g. a weather prediction institution that wants to decide which models have performed best by the end of each year. A sequential version of the MCS algorithm is provided that allows to compare and select the models sequentially over time incorporating the possibility of time-varying performances of the models. The approach is based on e-processes which allow for safe anytime-valid inference and have recently found great attention in the statistical literature.