Financial time series prediction using spiking neural networks

Area of Science:

• Computational Neuroscience
• Financial Forecasting
• Machine Learning

Background:

• Financial time series data are inherently non-stationary and noisy, posing significant challenges for traditional prediction models.
• Spiking Neural Networks (SNNs), particularly Polychronous Spiking Networks (PSNs), possess inherent temporal processing capabilities that may be advantageous for such data.

Purpose of the Study:

• To evaluate the efficacy of a Polychronous Spiking Network for financial time series prediction.
• To benchmark the PSN's performance against traditional neural networks and linear models.

Main Methods:

• A Polychronous Spiking Network was applied to predict three financial time series: IBM stock data, US/Euro exchange rate, and Brent crude oil prices.
• Performance was compared against a Multi-Layer Perceptron, a Dynamic Ridge Polynomial network, and a Linear Predictor Coefficients model.

Main Results:

• The Polychronous Spiking Network demonstrated favorable prediction results, including higher Annualised Return and lower prediction error for 5-step ahead predictions.
• Additional metrics like Maximum Drawdown and Signal-To-Noise ratio supported the superior performance of the PSN.

Conclusions:

• Polychronous Spiking Networks are applicable and effective for financial data forecasting.
• PSNs offer a potential advantage over traditional systems for predicting complex, non-stationary financial time series.