Implementation of Kriging in Forecasting Time Series Processes

This paper contains comparison of kriging technique implementation for time series modeling and forecasting. In the computational experiment kriging is compared to the standard methods (ARIMA/ETS). Basically, the concept of kriging was built in the task of multivariate function interpolation in the study of spatial distribution modeling of minerals and (when time is included in the models) geological processes. If one abstracts from the geological nature of the data, this approach can also be applied to modeling of an arbitrary multivariate quantity (constant or time-varying). The interest in this method is also related to the fact that the adjustable model belongs to a very simple class: linear or polynomial regression. In the presented research forecasting for some time interval in the future is made and procession of missing values is investigated (concentrated in one place or scattered over the data series). Combined models are built (the prediction is given as weighted combination of forecasts of several models). Each successive model is trained on the residuals of the previous models (it’s close to the bagging technique). The combination of ARIMA, ETS, and kriging models is studied.