Two Criteria for Time Series Models Comparison
Boštjan Vovk, Barbara Motnikar, Drago Čepar · European Consortium for Mathematics in Industry · 1992
Many methods for time series modelling have been proposed in literature. A general measure describing how a given model is suitable for representation of a given time series may be of interest in a broad field of applications. We suggest two criteria for time series models comparison. The first criterion estimates the degree to which a time series is explained by a given model. The criterion may be used for determining the optimal model obtained by the given method and for comparison of models obtained by different methods. We define dynamics of time series as mean square difference between neighbouring observations and a model activity criterion as the proportion of the explained time series dynamics. It will be shown that models with low value of model activity and high value of the explanatory criterion do not capture neither the long term components of the process nor the short term ones. Suggested criteria are compared with some other well-known criteria. Use of criteria is presented on the ARIMA models by some examples.