CAN NONLINEAR NOISE REDUCTION HELP NEURAL NETWORKS IN FINANCIAL FORECASTING

摘要

Generally, neural networks (NNs) learn and generalize directly from noisy data with the faith on ability to extract the underlying deterministic dynamics from the noisy data. However, it is well known that the model's generalization performance will be poor unless we prevent the model from over-learning. Given that most financial time series contain dynamic noise, it is necessary to reduce noise in the data with nonlinear methods before fitting the prediction models, In this study, we examine whether nonlinear noise reduction can help NNs in financial forecasting. Simple nonlinear noise reduction (SNL) and locally projective nonlinear noise reduction (LP) are applied to exchange rates time series, namely U.S. Dollar against the British Pound (GBP) and Japanese Yen (JPY). Both noisy and filtered data are used as input for NNs forecasting, in order to see whether nonlinear noise reduction could improve the predictions performance in term of Root of Mean Squared Error. Experiment results show that the two nonlinear noise reduction methods can improve the prediction performance of NNs. The improvement is not statistically significant in most cases based on the modified Diebold-Mariano test. Noise reduction methods work more effectively on JPY than on GBP. Because JPY is more volatile than GBP, that is, there are more noises in JPY than in GBP to be filtered. In the most cases, NNs outperform the random walk model. Perhaps we need more effective nonlinear noise reduction methods to improve prediction performance further. On the other hand, it indicates that NNs are particularly well appropriate to find underlying relationship in the environment characterized by complex, noisy, irrelevant or partial information.