A multiscale neural network learning paradigm for financial crisis forecasting

A financial crisis is typically a rare kind of an event, but it hurts sustainable economic development when it occurs. This study proposes a multiscale neural network learning paradigm to predict financial crisis events for early-warning purposes. In the proposed multiscale neural network learning paradigm, currency exchange rate, a typical financial indicator that usually reflects economic fluctuations, is first chosen. Then a Hilbert-EMD algorithm is applied to the currency exchange rate series. Using the Hilbert-EMD procedure, some intrinsic mode components (IMCs) of the currency exchange rate series, with different scales, can be obtained. Subsequently, the internal correlation structures of different IMCs are explored by a neural network model. Using the neural network weights, some important IMCs are selected as the final neural network inputs and some unimportant IMCs that are of little use in mapping from inputs to output are discarded. Using these selected IMCs, a neural network learning paradigm is used to predict future financial crisis events, based upon some historical data. For illustration purpose, the proposed multiscale neural network learning paradigm is applied to exchange rate data of two Asian countries to evaluate the state of financial crisis. Experimental results reveal that the proposed multiscale neural network learning paradigm can significantly improve the generalization performance relative to conventional neural networks. (C) 2009 Elsevier B.V. All rights reserved.