Quaternion Dendritic Neuron Model for Multivariate Financial Time Series Prediction

In prediction tasks, the single dendritic neuron models (DNMs) have achieved good results due to their inherent biological dendrite-like nonlinear calculation capabilities. Meanwhile, quaternion neural networks consisting of multi-layers of McCulloch-Pitts neurons have achieved remarkable achievements in spatial rotation, image processing, and multidimensional prediction. However, a single DNM has never been extended to quaternion domains and has not been applied to multivariate prediction tasks. In this work, we first generalize the real-valued DNM to the quaternion field. The performance of quaternion DNM (QDNM) is evaluated through several real-world multivariate financial time series prediction tasks. Also, the form of the forward phase of the neuron structure is analyzed comparatively. Experimental results demonstrate that the proposed QDNM achieves better results on diverse tasks than existing classical real-valued models and quaternion networks.