Interpretable approaches for financial time series forecasting

Abstract

Nowadays complex and non-linear real-world challenges have led to the development of advanced machine learning models. In most machine learning techniques, the ability to interpret the model’s predictions is a critical aspect. On the other hand, it is still difficult to understand if it is possible to provide a taxonomy of the various definitions of interpretability. In this thesis, a grey-box method, namely the Adaptive Neuro-Fuzzy Inference System with fractional Tikhonov regularization (ANFIS-T) was adopted to predict stock prices. Its performance was compared against white-box models, such as the autoregressive integrated moving average (ARIMA) and Naive, and the standard Adaptive Neuro-Fuzzy Inference System (ANFIS) model, with an emphasis on the interpretability factor. The root mean squared error (RMSE) and training time by the different approaches were compared. Overall, in this study, our investigation focuses on interpretability, aiming to offer an overview of all the definitions and related matters, and to discuss the ANFIS-T model’s performance for a forecasting problem, concluding that such performance is on par with white-box models. It also highlights the ANFIS-T model’s interpretability, offering helpful insights for situations involving decision-making where transparency and understandability are essential elements.