Many time series forecasters assume linearity in the data. Some also assume that whatever it is that is driving the outcome will continue to do so in the same way. Some researchers have acknowledged that the underlying processes are non-linear but claim that their method is adequately robust and that the error from this naive assumption is minimal.; Time series forecasting assumes that the near future will resemble the not too distant past. Research has shown that the performance of a forecast method is affected by the training method, and upon which data it was trained. Training is a process by which the parameters of the model are estimated. Various methods of organizing the data for this estimation process exist and this dissertation will study and compare the results of using rolling, moving, and weighted windows for this training process.; Some processes have changes in their causal factors, occurring over time. Some of these occur slowly while others occur quickly or are even in the form of system shocks. One or more of these influences can be occurring simultaneously. The weighted window gives more weight to the more recent or newest data and less to the older. As in the moving window, data beyond a certain age was eliminated from the training set. A variety of different values for the core and support widths was tried.; Nine economic data sets were used in this study to compare the performance of the three data window methods on observed forecast error. Both OLS autoregression and neural networks were employed as forecast tools. Forecasts on three data sets proved statistically different and better using weighted window training method. This reduction in error averaged almost 50%. Two more data sets had about a 15% reduction in error, but due to large variance within treatment were not proven different from the mean. No difference was proven statistically in the remaining four data sets. No difference between using any of the three (3) window methods with regression was proven.
展开▼
