Mårten Waern och Fredrik Svedberg present their master's thesis

Analyzing Foreign Exchange Rates: A Neural Network Approach

Abstract

The problem of success fully predicting foreign exchange rates stands
unsolved to this day. Non-linear time series models combined with
computer supervised data mining comprise a potential way to go about
it, given that the problem is possible to model. We have opted for a
neural network structure coevolving a Marquart-Levenberg
backpropagated optimization algorithm with an evolutionary time
dependent optimization over a whole population of networks. Our
results vary, consistently outperforming the random walk model in some
cases while not in others. Overall the results are encouraging.