Performance Evaluation of Artificial Neural Networks in the Foreign Exchange Market

This thesis examines the performance of artificial neural networks in the foreign exchange market. The thesis is restricted to comprise two types of network architectures: feedforward and probabilistic neural networks, respectively. The networks’ capabilities are evaluated in a trading simulation, where predictions of exchange rate log-returns are backtested using historical data. All G10 currency pairs are considered, 45 in total. The results presented indicate that although several networks generate substantial returns, the average performance is rather modest. The foreign exchange market indeed appears efficient. Acknowledgements The author would like to thank Nils Rosendahl and Alexander Wojt at Nordea Markets in Stockholm for the opportunity given to write this thesis, and for the many rewarding discussions regarding the architecture of the artificial neural networks presented in the following. Naturally, the author assumes full responsibility for any errors present in the paper.