Knowledge-based selection and combination of forecasting methods

Abstract

Forecasting techniques are used extensively in a variety of business (and other) applications for the analysis and prediction of various factors of interest e.g. future market demand for a specific product. However, the selection and use of the appropriate models is a non-trivial problem given the broad range of qualitative and quantitative factors that have to be considered. Several researchers have investigated the use of knowledge based "expert" systems to aid the selection of specific techniques. Although such systems may provide a moderately good list of potential methods, they tend to be deficient in selecting an efficient "cover" of complementary forecasting models. Some authors have argued the need to distinguish· "macro" and "micro" representations of knowledge in systems to automate model selection. The macro-level knowledge would provide semantic or contextual knowledge about the domain of application. Such knowledge is needed to provide the deeper explanation required if the system is to be capable of use by statistical novices. It is also necessary if any assisted interpretation of the results is anticipated. This paper will review research on the use of frames to represent such macro knowledge and will discuss the use of a technique called abductive inference to combine several methods to select the best composite group of f orecasting techniques.