Increasing interpretability using a fuzzy-embedded recurrent neural network (FE-RNN) with its application in stock ETF trading

• Main Author: Tan, James Chee Min
• Other Authors: Quek Hiok Chai
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/148790
• Institution: Nanyang Technological University

Deep learning has been a recent breakthrough that has enabled predictions and modelling to be very accurate. These predictions and modelling tools were once used to help us understand our data and serve as a tool to make a judgement. However, the vast improvements in these deep learning structures have enabled them to perform decision-making independently. Many decisions made by such deep learning models have been tested to be much better at performing their task than when we used these models merely as a tool. The problem with deep structures is that they lack the interpretability of conventional modelling techniques such as a traditional fuzzy inference system. This paper proposes a fuzzy-embedded deep structure, the fuzzy-embedded recurrent neural network (FE-RNN). FE-RNN uses a one-pass DIC clustering method to form fuzzy membership values to feed into the recurrent neural network. The structure utilises pseudo-outer product rule generation to interpret the embedded recurrent neural network. Finally, the model's crisp output can be obtained through centre-of-gravity defuzzification. As both the deep structure and the fuzzy structure share a common input and output linguistic, we are able to associate the inference process of the RNN with fuzzy rules. The fuzzy IF-THEN rules help us interpret the inference process of the FE-RNN. The performance of FE-RNN is evaluated and compared against the vanilla RNN and other fuzzy neural network structures through benchmark experiments in the Mackey-Glass dataset, Nakanishi datasets and price forecasting for various indices such as the S&P500 & DJI. They produce good results in benchmark experiments but suffer in the Nakanishi dataset, where the training data is sparse. The learning process and inference process is then visualised to associate the rule nodes with the deep recurrent nodes in the RNN. Lastly, its prediction is used in a GA-fMACDH trading system that has found to outperform the buy and hold strategy in most of the ETFs experimented with in the backtesting period.