High Predictive Performance of Dynamic Neural Network Models for Forecasting Financial Time Series

The study presents high predictive performance of dynamic neural network models for noisy time series data; explicitly, forecasting the financial time series from the stock market. Several dynamic neural networks with different architecture models are implemented for forecasting stock market prices and oil prices. A comparative analysis of eight architectures of dynamic neural network models was carried out and presented. The study has explained the techniques used in the study involving the processing of data, management of noisy data, and transformations stationary time series. Experimental testing used in this work includes mean square error, and mean absolute percentage error to evaluate forecast accuracy. The results depicted that the different structures of the dynamic neural network models can be successfully used for the prediction of nonstationary financial signals, which is considered very challenging since the signals suffer from noise and volatility. The nonlinear autoregressive neural network with exogenous inputs (NARX) does considerably better than other network models as the accuracy of the comparative evaluation achieves a better performance in terms of profit return. In non-stationary signals, Long short term memory results are considered the best on mean square error, and mean absolute percentage error.