Retno Aulia Vinarti, Arif Djunaidy : FOREIGN EXCHANGE FORECASTING USING A GENETIC RECURRENT NEURAL NETWORK

Retno Aulia Vinarti S.Kom., M.Kom.
Prof. Ir. Arif Djunaidy M.Sc., Ph.D.

Year

2011

Published in

ISICO 2011

External link

-

Type

-

Keywords

forecasting of foreign currency exchange recurrent neural networks genetic algorithm minimum error maximum profit


Abstract

The use of recurrent neural networks (RNNs) for foreign exchange forecast gives pro¬mis¬¬¬¬¬ing results because of their excellent perfor¬man¬ce in minimizing the error of the model, parti¬cu¬larly for the data having chaotic behavior. How¬ever, in practice, the use of a fore¬casting method that is merely capable of mini¬mizing the model’s error but lacking of producing an optimal forecasting profit is impractical for most traders. To overcome such a drawback, in this study the modified RNN is integrated with a genetic algorithm (GA) in or-der to provide a forecasting model that is capable of both minimizing the error and maximizing the profit of the model. Modified RNN is placed on the performance function of neural network. Formerly, a default performance function of RNN is only error minimizing. In this research work, we build a new performance function, which is maximizing profit of Foreign Exchange Trading. The re¬sult¬ing forecasting mo¬del was tested using a daily time series obtained from http://www.tradingblox.com for a period rang¬¬ing from January 4, 1999 until October 20, 2010. Experi¬men¬tal results—using three types of foreign exc¬ha¬nge that are fre¬qu¬ent¬ly used in practice i.e., USD/JPY, EUR/USD, and GBP/USD—show that the forecasting model that both minimizing the error and maxi¬mizing the profit of the fore¬casting model can be achieved using a modified RNN employing a pair of perfor¬mance functions that are optimized using GA.