Evolving data-driven Interpretable fuzzy deep neural network (IFDNN) with applications in algorithmic finance

Abstract

Deep learning has been a fast-growing field in computer science. It is a state-of-the- art machine learning approach that has shown promising results in many areas. Its ability to learn intricate and complex structures within large amounts of data makes it powerful in learning non-linear patterns in data. Thus, with a well-trained model, the quality of predictions are highly accurate. However, the issue with deep learning models is that they lack interpretability. Despite having highly accurate predictions, users are not able to understand the reasoning behind the predictions. This might not be a problem in certain fields, but in tasks involving degrees of human interference, it would be desirable to understand the inference process in the deep structure, enabling more informed decisions to be made. This dissertation proposes a data-driven Interpretable Fuzzy Deep Neural Network model (IFDNN) that provides insight into neural network inferences. It will be designed such that it is able to handle concept drift, which are changes in the underlying distribution of data across different periods. It is a common problem in predictive modelling and if left unaddressed, results in poor predictions from inadequate learning. In particular, we will be deploying IFDNN on financial market data, which contains concept drift. Subsequently, we utilise IFDNN’s forecasts for multiple look-ahead timesteps to accurately detect trend reversals. This enables us to improve traditional momentum indicators, with its effectiveness evaluated in trading and portfolio rebalancing experiments.