MIXED AND MEMORY FADING FORECASTING MODELS
Abstract (summary)
This study examines the problem of forecasting time series related processes. Several forecasting procedures are examined and modifications to these methods are proposed in the paper.
Much interest has been shown in the past few years in the analysis of various time series related forecasting procedures. The first part of the study is devoted to a survey of papers on the topic. A summary of the more relevant research articles is given over the past several years.
The second part of the study includes an examination of the simple moving averages, exponential, and mixed exponential-simple moving averages models. The mathematical and statistical properties of the models are examined and motivation is given for the combining of forecasting techniques. The results of an empirical study using the mixed forecasting model are given.
The importance of minimizing forecast error is brought out in the next part of the work. The significance of "adaptive forecasting" in error reduction is emphasized here and ways to make the mixed exponential-simple moving averages model "adaptive" are studied. An examination of the variance of forecast error is discussed, and its use in the development of prediction intervals for future estimates of data values is analyzed.
The last part of the work is devoted to the analysis of the autoregressive, moving average and mixed autoregressive-moving average models. The mathematical forms of the models are set forth and examples of special cases of the models are given. These models are very powerful forecasting procedures, but have several disadvantages which are mentioned in the thesis.
One of the disadvantages of the models noted above, is that a relatively large number of data values is needed for the construction of a model. In some cases the older data values may be of little use in the forecasting of future data values. A method of discounting the significance of the older data values is explored here. A comparison of forecast error is made between the regular and discounted methods using three actual data sets.