Evolving neuro-fuzzy system for portfolio management

Abstract

Recently, Explainable Artificial Intelligence (XAI) has been on the rise. More companies are opting for models which are both highly accurate and highly interpretable. Although, traditional Neural Networks produce great results, they are black-box models which lack in interpretability. Thus, in this project, a partially online evolving density-based fuzzy convolutional neural network (EDFCNN) was developed to predict stock prices. The fuzzification of the deep learning network was a key component of the project to ensure that the model maintains high standards of interpretability. It provides the linguistic basis on which we can understand each prediction made by EDFCNN. With the very accurate predictions made by EDFCNN, we then developed an improved financial indicator: the Predicted Moving Average Convergence Divergence (MACD). We used an online density-based clustering algorithm, DBSTREAM, to generate the fuzzy input space. The fuzzy inputs were then passed into a five-layer feed-forward Convolutional Neural Network which predicted the fuzzy output values. The principle of memory decay was applied to update the weights of the rules in the rule base to mimic the way humans retain information. EDFCNN displayed superb performance (R2 > 0.99) consistently on various types of financial securities and market conditions. EDFCNN was able to improve on the Vanilla MACD Indicator by reducing the RMSE to the Perfect MACD Indicator by >30% in all test cases. EDFCNN can be used as a trading instrument to predict future stock prices and generate a forecasted MACD indicator. This will allow long term traders to make more informed trades to maximise profits. Additionally, when the Predicted MACD was combined in an automated portfolio rebalancing system, investors can be exposed to even more upside – in terms of risk mitigation and profits.