A New Description Length Criterion for Assessing Local Complexity in Neural Networks and Other Nonlinear Statistical Models with Applications to Financial Time Series Forecasting
by
David Edelman
University of Wollongong

Various types of Description Length Criteria have recently been proposed for generalising traditional Likelihood-Based estimation Methods to include Model Complexity Costs, with the principles being generally well-received, but with no clear consensus as to a single most 'sensible' (Canonical) way to apply them in practice. It is shown in the present paper that in the case of nonlinear statistical models, there exists a natural measure of Model Complexity which is not necessarily constant throughout a given parameter space, which (arguably) should be taken into account when fitting or 'training' such a model. It is first argued, then demonstrated with real Financial data, that the resulting criterion may be used to help avoid overfitting and to improve generalisation error.

Date received: November 12, 2001