CMStatistics 2015: Start Registration
View Submission - CMStatistics
B1682
Topic: Contributions on independence and distance-based methods Title: Predictive model choice in time series analysis under stationarity and non-stationarity Authors:  Piotr Fryzlewicz - London School of Economics (United Kingdom)
Tobias Kley - University of Bristol (United Kingdom) [presenting]
Philip Preuss - (Germany)
Abstract: In statistical research there usually exists a choice between structurally simpler or more complex models. We argue that, even if a more complex model were true, a simple one may be advantageous to work with under parameter uncertainty. We present an alternative model choice methodology for time series analysis, where one of two competing approaches is chosen based on its empirical finite-sample performance with respect to a certain task. A rigorous, theoretical analysis of the procedure is provided in the framework of choice between stationarity and local stationarity when the task is to forecast. We state conditions that imply when it is preferable to base the forecasts on the more volatile time-varying estimates and when it is advantageous to forecast as if the data were from a stationary process, even though it is not. We also consider different frameworks, as for example choosing between linear and non-linear time series models, and provide the results of an extensive simulation study as well as an empirical example.