An interpretable Neural Fuzzy Hammerstein-Wiener network for stock price prediction

Abstract

An interpretable regression model is proposed in this paper for stock price prediction. Conventional offline neuro-fuzzy systems are only able to generate implications based on fuzzy rules induced during training, which requires the training data to be able to adequately represent all system behaviors. However, the distributions of test and training data could be significantly different, e.g., due to drastic data shifts. We address this problem through a novel approach that integrates a neuro-fuzzy system with the Hammerstein-Wiener model forming an indivisible five-layer network, where the implication of the neuro-fuzzy system is realized by the linear dynamic computation of the Hammerstein-Wiener model. The input and output nonlinearities of the Hammerstein-Wiener model are replaced by the nonlinear fuzzification and defuzzification processes of the fuzzy system so that the fuzzy linguistic rules, induced from the linear dynamic computation, can be used to interpret the inference processes. The effectiveness of the proposed model is evaluated on three financial stock datasets. Experimental results showed that the proposed Neural Fuzzy Hammerstein-Wiener (NFHW) outperforms other neuro-fuzzy systems and the conventional Hammerstein-Wiener model on these three datasets.