Self-evolving neural fuzzy system with application in portfolio management

Abstract

Deep learning has gained popularity over the recent years and have shown success in solving complex problems in various disciplines with a high prediction accuracy. However, the black-box nature of many deep learning models has resulted in experts facing difficulties in explaining the mappings between the neural network layers and how the predictions are derived, which may be crucial when it comes to deploying the model in high stakes decision making environment such as the healthcare or financial sector. This paper proposes a Self-Evolving Neural Fuzzy System which aims to improve the interpretability of the deep neural network by integrating fuzzy system with the neural network. It is coupled with online modelling technique, such as an online clustering algorithm for fuzzy membership generation to dynamically adapt to new patterns in data stream. This ensures that the model maintains a high accuracy by learning the underlying characteristics of the data incrementally, even in the event of a concept drift. The proposed system will be tested on financial market data that is known to be volatile in nature, to evaluate its prediction accuracy of future stock prices. Subsequently, the forecasted stock prices are incorporated into a lagging technical indicator such as MACD, to reduce the time lag present such that trend-reversal signals can be better identified to determine the optimal trading strategy that gives the best return for our portfolio. The proposed trading strategy is then benchmarked against the “Buy and Hold” strategy and vanilla MACD strategy and the performance of the portfolio is evaluated under different market conditions. SENFS shows promising results, achieving a high R^2 score above 0.9 for predictions up to 5 days lookahead and the forecasted MACD outperformed the other trading strategies during bear and volatile market, only losing to the “Buy and Hold” strategy in the bull market by a small margin.