Evaluation of Multivariate Recurrent Neural Networks Ability to Predict Changes in USD/SEK Exchange Rate

Sammanfattning

In recent years, recurrent neural networks have gained a lot of popularity when forecasting financial time series. In this thesis we evaluate how well recurrent models can predict next days return of the USD/SEK exchange rate. The thesis also investigates if models incorporating exogenous features outperform single feature models. Three networks types are considered, simple RNN, LSTM and GRU. The results imply that multi-feature models perform significantly better than single feature models, and the most important exogenous features are the yields of Swedish and US government bonds with a duration of 0.5 to 10 years. GRU is the network type achieving the highest performance across almost all performance metrics on three different test/validation sets. With an average directional accuracy of 0.54, across 50 runs on the validation set. Which when translated to a long-short trading strategy correlates to an average annualized return of 23,46%. LSTM and simple RNN achieved annualized returns of 19.39% and 16.05% respectively.