| Abstract [eng] |
This thesis investigates whether a single-hidden-layer Feedforward Neural Network (FFNN), based on Cybenko’s Universal Approximation Theorem, can achieve practically viable directional pre- diction accuracy for the EUR/USD closing price when trained on a pre-filtered dataset of histori- cally similar market patterns. Over twenty years of daily EUR/USD data were converted to percentage returns and orga- nized into five-day sliding window feature vectors. A Euclidean distance algorithm filtered similar patterns using threshold τ = 1.4, yielding 145 learning-phase and 11 test-phase patterns. FFNNs were trained on this filtered subset using sigmoid and ReLU activation functions, evaluated across 100 random seeds, and extended with a K-Nearest Neighbours consensus ensemble. The sigmoid model ranged from 27.27% to 81.82% test accuracy, while the ReLU model ranged from 9.09% to 81.82%, with both 100-seed medians stabilizing at 54.55%. The best single-seed ReLU configuration achieved 72.73% test accuracy, correctly classifying all negative outcomes. The ReLU-KNN consensus ensemble produced an agreement rate of 81.82% with a consensus accuracy of 55.56%. The study concludes that noise reduction through Euclidean distance filtering is a decisive factor in achieving reliable directional predictions, enabling a minimal neural network architecture to perform meaningfully on unseen financial data. |
