Neural network-based decision support systems have been applied to myriad forecasting problems. Donaldson and Kamstra (1996) consider the use of neural networks in combining time series forecasts. They conclude that the ability of neural networks to account for non-linear relationships provides fora possible superiorityof the neural network combination approach over traditional linear combination procedures. Other noteworthy examples include the application of network-based systems to auditing (Lenard, Alam and Madey, 1995), bankruptcy forecasting (Fletcher and Goss, 1993), statistical classi®cation problems (Markham and Ragsdale, 1995) and predicting student grade point averages (Gorr, Nagin and Szczypula, 1994). Lachtermacher and Fuller (1995) explore theuse of the BoxandJenkinsmethodologyas ameans for reducingthe data requirements necessary to train a backpropagation neural network. There has been considerable
Copyright #1999 John Wiley & Sons, Ltd. J. Forecast. 18, 167±180 (1999)
attention directed to the use of neural networks as an alternative to quantitative forecasting and decision models (see Hill et al., 1994 for a surveyof this literature and Hill, O'Connor and Remus, 1996 for a comparison between neural networks and statistical time series methods). It is important to note that the focus of this paper is not to consider neural networks as an alternative forecasting methodology. In contrast, the intelligent forecasting system, utilizing aneural network structure in the selection phase, provides a guide to the problem of forecasting in general and model selection in particular, and develops a framework for directly incorporating time series characteristics into the forecasting system. In this context, the intelligent forecasting system serves as an approach for reducing the set of potential forecasting models under consideration and as such, may provide guidance to the practitioner in the selection of forecasting models.
This paper develops a decision support system that combines calibration, model selection and forecast generation into an intelligent forecasting system. As illustrated in Figure 1, for a given data set a model is selected based on a number of criteria and these criteria vary across user and organization. As the literature indicates, familiarity with a forecasting method is the ®rst step towards usage. For the case of multiple forecasts, such as products within a product line, a methodology may be selected for an entire group of data. The model selection phase is a timeconsuming and arduous task and the inertia to retain the methodology and recalibrate the existing model at selected intervals of time is strong. In the intelligent forecasting system, rather than recalibrating the model at selected intervals, the practitioner is recalibrating the model selection process and the model parameters, resulting in a model designed to provide quantitative forecasts accurately. In the decision support system the focus becomes using the most recent time series information to both select and estimate the quantitative model.
The paper is outlined as follows. First, the architecture of the intelligent model selection and forecasting system is discussed. The following section discusses the system architecture, with a description of the neural network architecture and a discussion of the results of the training and testing of the system. The ®nal section provides the conclusions and some directions for future research.
The architecture of the neural network based model selection and forecasting system utilizes a three phase approach as shown in Figure 2. The initial phase (Phase 1) considers characteristics of the time series. These distinguishing features of the time series include sample size, data frequencies and related data issues (Table I) that often impact heavily on forecast performance.
Table I. Names and types of input neurons
Уважаемый посетитель!
Чтобы распечатать файл, скачайте его (в формате Word).
Ссылка на скачивание - внизу страницы.