Effective Sampling and Windowingfor an Artificial Neural Network Model Used in Currency Exchange Rate Forecasting

Sammanfattning: Financial forecasting is a field of great interest in academia and economy. The subfield of exchangerate prediction is of considerable value to practically every entity operating within the financialmarket. Ranging from private hedgers, speculators or arbitrageurs to entire financial institutions suchas international banks or insurance companies, the ability to predict exchange rate movementsprovides major benefits for organizations in contact with these. A great multitude of research has beenconducted to construct methods to aid firms and investors to better anticipate on potentialdevelopments in the foreign exchange market. Much of the research has been focusing on a promisingprediction model within computational intelligence developed in recent decades, namely the ArtificialNeural Network (ANN). However, a review of existing literature suggests that the time step, predictionhorizon and window size have not been of central essence. Hence, this paper attempts to provide amore formal analysis of the actual impact of the three mentioned parameters on the prediction resultsof ANNs. Through literature studies, modeling and experimentation it is found that no specificcombination of time step, prediction horizon and window size results in more exact forecasts, but thatcertain combinations of the three parameters generally result in superior performance.