A survey by Wong et al. (1995) confirms that the vast majority of business-related neural network (NN) applications use the multilayer perceptron (MLP) model with the backpropagation algorithm employed to carry out training. The authors prefer to refer to the network in terms of the label MLP because it is important to distinguish between the topology of the network and the algorithm employed to train it. As regards using NN methods for forecasting the same standard approach has also been employed. Previous work by the authors suggests, however, that there is at least scope for doubting what has become an orthodoxy and has also considered certain approaches to diagnostic testing. Work has also been carried out on the precise details of the approximations provided by network models. This has looked at power series equivalents of the standard MLP model and has also considered a network's capacity to mimic or approximate certain functional forms. The research in question, described in more detail below was carried out without reference to specific applications: it also leaves scope for extended study. The paper considers the various strands of the work more specifically with reference to business forecasting. It also considers how the known properties of network models and techniques may need to be reassessed in a time series context.
[1]
Bo K. Wong,et al.
A bibliography of neural network business applications research: 1988–September 1994
,
1995
.
[2]
Andrew R. Barron,et al.
Universal approximation bounds for superpositions of a sigmoidal function
,
1993,
IEEE Trans. Inf. Theory.
[3]
William Remus,et al.
Neural Network Models for Time Series Forecasts
,
1996
.
[4]
M. B. Priestley,et al.
Non-linear and non-stationary time series analysis
,
1990
.
[5]
Duarte Trigueiros,et al.
Neural networks and empirical research in accounting
,
1996
.
[6]
B. Curry,et al.
Neural networks: a need for caution
,
1997
.
[7]
Wilpen L. Gorr,et al.
Editorial: Research prospective on neural network forecasting
,
1994
.