Predicting Forex Currency Fluctuations Using a Novel Bio-inspired Modular Neural Network

This thesis explores the intricate interplay of rational choice theory (RCT), brain modularity, and artificial neural networks (ANNs) for modelling and forecasting hourly rate fluctuations in the foreign exchange (Forex) market. While RCT traditionally models human decision-making by emphasising self-interest and rational choices, this study extends its scope to encompass emotions, recognising their significant impact on investor decisions. Recent advances in neuro- science, particularly in understanding the cognitive and emotional processes associated with decision-making, have inspired computational methods to emulate these processes. ANNs, in particular, have shown promise in simulating neuroscience findings and translating them into effective models for financial market dynamics.

However, their monolithic architectures of ANNs, characterised by fixed struc- tures, pose challenges in adaptability and flexibility when faced with data perturbations, limiting overall performance. To address these limitations, this thesis proposes a Modular Convolutional orthogonal Recurrent Neural Net- work with Monte Carlo dropout-ANN (MCoRNNMCD-ANN) inspired by recent neuroscience findings.

A comprehensive literature review contextualises the challenges associated with monolithic architectures, leading to the identification of neural network structures that could enhance predictions of Forex price fluctuations, such as in the most prominently traded currencies, the EUR/GBP pairing. The proposed MCoRNNMCD-ANN is thoroughly evaluated through a detailed comparative analysis against state-of-the-art techniques, such as BiCuDNNL- STM, CNN–LSTM, LSTM–GRU, CLSTM, and ensemble modelling and single- monolithic CNN and RNN models. Results indicate that the MCoRNNMCD- ANN outperforms competitors. For instance, reducing prediction errors in test sets from 19.70% to an impressive 195.51%, measured by objective evaluation metrics like a mean square error.

This innovative neurobiologically-inspired model not only capitalises on modularity but also integrates partial transfer learning to improve forecasting ac- curacy in anticipating Forex price fluctuations when less data occurs in the EUR/USD currency pair. The proposed bio-inspired modular approach, incorporating transfer learning in a similar task, brings advantages such as robust forecasts and enhanced generalisation performance, especially valuable in domains where prior knowledge guides modular learning processes. The proposed model presents a promising avenue for advancing predictive modelling in Forex predictions by incorporating transfer learning principles.