MODEL DEEP LEARNING LSTM DAN GRU UNTUK PREDIKSI VALUTA ASING

Abstract

Currency exchange rates are one of the most important variables in macroeconomics because changes in exchange rates can affect economic stability and activity. The uncertainty of exchange rates is an important problem in finance, so accurate predictions are needed to design strategic steps in dealing with foreign exchange fluctuations. This research aims to explain the process of implementing the LSTM and GRU deep learning models in predicting foreign exchange against the Rupiah and US Dollar exchange rates and analyzing the evaluation results and prediction results of the LSTM and GRU models. The data used is the daily closing price of the USD/IDR exchange rate for the period August 29, 2014, to August 28, 2024, obtained from the global financial platform investing.com, totaling 2,554 data points. Predictions are made for a one-step forward horizon (t+1) with a chronological data division scheme of 80% as training data and 20% as testing data. Before the model training process, the data is normalized using the Z-score normalization method to improve data stability against outliers and aid the model learning process. The results showed that the LSTM and GRU models were able to learn historical data patterns well during the training and testing processes. During testing, the LSTM model produced an MAE of 0,133; RMSE of 0,159; and  of 0.932. After denormalization, the LSTM model has a mean prediction error (MAE) of approximately Rp65 and a RMSE of approximately Rp89 relative to the actual value. Meanwhile, the GRU model produced an MAE of 0.102; RMSE of 0.124; and  of 0.958. The denormalization results show that the GRU model has an average prediction error (MAE) of approximately Rp54 and an RMSE of approximately Rp70 relative to the actual value. The evaluation results show that the GRU model performs better than the LSTM model because it produces a smaller error value and the  value that is closest to 1.