Predicting fluctuations in foreign exchange rates

This paper assesses the viability of predicting fluctuations in the foreign exchange markets. It investigates the new Weka filter that evolves the input space of the decision system. The new Weka filter only enhances one file and so falls short of the requirements for this project. Subsequent experiments used an earlier version that kept the training and testing data separate but enhanced both. The test data was taken from an earlier project that used the various techniques available in the Weka library and so the hypothesis tested was that the new system would give improvements. Various systems were constructed that simplified the execution of multiple tests. There are important factors that need to be taken into account when conducting learning schemes on continuous financial data. A price history model was constructed, which shows that short period financial predictions are very difficult to predict and highly volatile. The system was then tested on the new Weka filter based on genetic algorithms with a multiple price point model. This was compared against a prior publication using ensemble learning but only using the previously existing Weka library. The GA based system, which enhances the input space, and was the foundation of the new Weka filter but took into account the need for separate training and test data was used in the final tests. Results indicate that in an ensemble combination, this technique attains a higher accuracy than the earlier ensemble based learning system with a confidence of 97%.