Is It Possible To Use Intelligent Systems To Design A Profitable Foreign Exchange Trading Agent?

In this paper, a trading agent is developed using a basket of intelligent systems with the goal of trading the GBPUSD currency pair profitably in the Foreign Exchange market. The basket of intelligent system consists of two regression models: a radial basis neural network and a TSK-fuzzy inference system; and three classification models: k-nearest neighbour, support vector machine and a decision tree. The trading strategy combines the predictions of each model using a Kalman-type filter to arrive at a final consensus of 'price increase' or 'price decrease' for the next period; it then buys and sells accordingly. The inputs to the models are in the form of 19 technical indicators which are mathematical formulae derived from the past price data of the currency pair. Since there is little consensus about the timing and selection of the best indicators to use in the market, a genetic algorithm aims to optimize the number and input type for each model. Once the best indicators have been chosen, the system simulates trading at 15-minute, 6-hour and daily intervals over the 16-month period from January 1 2008 to April 26 2009. In order to be as accurate to real life conditions as possible, the system is measured on overall profitability defined in percentage in points (PIPs), total drawdown and compared with a random walk benchmark. The trading agent outperforms the random walk model in all simulations, however, it only realizes a positive overall profit when trading at daily intervals. Although the random walk benchmark is beaten quite substantially each time, the system does not consistently create profit at high trading frequency and therefore the Efficient Market Hypothesis (EMH) cannot be fully rejected. The fact that the model is profitable at lower frequency suggests some evidence against the EMH but further testing is necessary to arrive at more concrete results