The use of neural networks in the combining of time series forecasts with differential penalty costs

TR Number
Date
1993-10-15
Journal Title
Journal ISSN
Volume Title
Publisher
Virginia Tech
Abstract

The need for accurate forecasting and its potential benefits are well established in the literature. Virtually all individuals and organizations have at one time or another made decisions based on forecasts of future events. This widespread need for accurate predictions has resulted in considerable growth in the science of forecasting. To a large degree, practitioners are heavily dependent on academicians for generating new and improved forecasting techniques.

In response to an increasingly dynamic environment, diverse and complex forecasting methods have been proposed to more accurately predict future events. These methods, which focus on the different characteristics of historical data, have ranged in complexity from simplistic to very sophisticated mathematical computations requiring a high level of expertise. By combining individual techniques to form composite forecasts in order to improve on the forecasting accuracy, researchers have taken advantage of the various strengths of these techniques. A number of combining methods have proven to yield better forecasts than individual methods, with the complexity of the various combining methods ranging from a simple average to quite complex weighting schemes.

The focus of this study is to examine the usefulness of neural networks in composite forecasting. Emphasis is placed on the effectiveness of two neural networks (i.e., a backpropagation neural network and a modular neural network) relative to three traditional composite models (i.e., a simple average, a constrained mathematical programming model, and an unconstrained mathematical programming model) in the presence of four penalty cost functions for forecasting errors.

Specifically, the overall objective of this study is to compare the shortterm predictive ability of each of the five composite forecasting techniques on various first-order autoregressive models, taking into account penalty cost functions representing four different situations. The results of this research suggest that in the vast majority of scenarios examined in this study, the neural network model clearly outperformed the other composite models.

Description
Keywords
Citation