Case Study: Financial Time Series Prediction Using Spiking Neural Networks
Most financial data are non-stationary by default, in that the statistical properties of the data changes over time. These changes are a result of various business and economic cycles, such as the high demand for air travel in the summer months effecting exchange rates and fuel prices. Thus financial series, such as stock market fluctuations, are a result of complex interactions and are affected by many highly correlated economic, political and even psychological factors.
The prediction of financial time series is a difficult problem since it depends not only on known, but also on unknown factors, and frequently data used for the prediction can also be noisy, uncertain and/or incomplete.
Traditional methods for financial time series forecasting are based around standards statistical approaches and machine intelligence techniques such as support vector machine and fuzzy logic. These achieved limited success due to the highly nonlinear nature of most of the financial time series or tended to be focused toward specific application scenarios.
Neural networks are believed to have great potential due to their predictive ability, adaptability to different domains and robust behavioural characteristics in uncertain environments. Multi-Layer Perceptrons (MLPs) and recurrent Neural networks have been successfully applied to a broad class of financial markets predictions and for some cases Higher-Order Neural Networks also proved appropriate.
More recently, Spiking Neural Networks (SNNs) have demonstrated the power of this type of neural networks in solving difficult problems in complex, ‘informationally messy’ and dynamic environments, and hence represented the main approach used by this project for Financial Time-series prediction.
We have applied a Polychronous Spiking Network (PSN) to solve non-stationary financial data prediction problems in order to exploit the temporal characteristics of the spiking neural model in an appropriate way. Our spiking network model adopted the Izhikevich neural architecture using axonal delays encoding the information such that its temporal aspects were preserved.
Despite the fact that Time-series prediction using machine intelligence techniques has been researched and investigated for a long time, the practical prediction of financial time series continues to attract considerable interest from both research and business communities due to the potential for large financial gain associated with even very small efficiency gain margins in terms of the prediction accuracy.
Our experiments showed that the SNN exhibits favourable and promising prediction performance compared to other types of neural networks. Our Spiking Network outperformed Multi-Layer Perceptrons (MLP) and a Dynamic Ridge Polynomial Neural Network (DRPNN), when solving the three financial datasets of IBM stock data, the US/Euro exchange rate and the price of Brent crude oil, using the key financial measure of Annualised Return (AR) and the Mean Square Error.
From an algorithmic viewpoint, the results suggest that explicit engineering of the temporal aspects of the financial data into a suitable spiking neural network (SNN) ‘lends itself better’ to time series analysis than traditional rate encoded neural networks or the SNN paradigms derived from the rate encoded networks.