Implementasi Multilayer Perceptron Pada Jaringan Saraf Tiruan Untuk Memprediksi Nilai Valuta Asing

Abstract

Abstract. In the context of FOREX investment, the fluctuation of currency becomes a common thing in which movement is greatly influenced by supply and demand. If the demand is higher, the price will increase and conversely, if the supply is higher, the price will go downward. There is a principle that the behavior of price patterns will repeat randomly and make unpredictable movement of FOREX. These patterns of currency fluctuation have deceived many investors and brought losses and even capital failure. Basically, the value of foreign exchange belongs to the data of time series and Multilayer Perceptron is very suitable to process data of time series as it is often used to make prediction. Therefore, this research aimed at implementing Multilayer Perceptron in the artificial nerve network for predicting the value of foreign exchange on the available resources using the attributes of open, high, low, and close. To process the data from the existing attributes, there must be initialization first in X1 (open), X2 (high), and X3 (low) as the inputs and Y (close) as the data target, and then they were normalized so as to calculate sigmoid. The increasing number of epoch does not guarantee that the errors will be smaller. On the contrary, perhaps, the error value will increase. The best result of training occurred by epoch 200 and learning rate 3 within the smallest values of MSE 281.02518, MAD 13.168, and deviation standard 10.294.