Dynamic hybrid fuzzy deep neural networks for practical securities trading

Abstract

The current epoch of evolutionary deep neural networks has reached unprecedented scales of depth and size. Possessing unmatched learning capabilities which far exceed any existing forms of computing, they are the forefront of innovation in almost every field. Yet, one area of deep neural networks still under-explored is their black box nature. The wave of adoption of deep neural networks in modern machine learning applications has continued to leave the interpretability of such networks an open question. As more intelligent systems start to leave human decision-making out of the loop, the possibility of unaccountable decision-making left in the hands of AI machina continues to be a growing risk in the Information Age. Machine learning in finance is not a new concept, spanning decades of research at attempts to crack the code of commodity markets. As accountability to stakeholders and clients is one of the most important facets of any financial product and service, work must be done to improve how the inputs and outputs of neural networks can made comprehensible to human agents. Much of past work has implemented fuzzy logic into practical interpretable inference systems. This paper will examine the theoretical and practical implementation of a hybrid Takagi-Sugeno-Kang fuzzy neural network in the context of portfolio trading.