Eureka! Institutional Repository

This dissertation is concerned with foreign exchange rate analysis using deep learning machines. Deep learning is a rather new neural learning technique that uses multiple hidden layers in order to capture the correlations between the data. Specifically, experiments were conducted on the direction of the US dollar – Euro exchange rate using a model based on a deep belief network. A wide selection of economic variables was used as an input. This model is comprised of stacked RBMs, whose weights were used to initialize a Multi Layer Perceptron model with a similar structure, trained with Conjugate Gradient optimization. In order to improve the generalization efficiency of the model, an investigation was conducted regarding the selection of the best features based on the information gain. The results showed that we can achieve considerably improved accuracy, compared to previous attempts on foreign exchange rate prediction.