Improving the Performance of Long Short Term Memory and Gated Recurrent Unit for Predicting the Composite Stock Price Index

Investing in stocks has become a popular choice in Indonesia, as evidenced by the significant increase in the number of retail investors. However, uninformed investment decisions, particularly by novice investors who tend to follow trends without thorough analysis, can result in significant losses. Therefore, utilizing predictions of the composite stock price index to design a portfolio strategy can serve as a basis for making informed investment decisions. In previous studies, predictions were made using basic features such as opening price, closing price, highest price, lowest price, and volume. This research proposes an improvement of LSTM and GRU methods for predicting stock prices by employing Grid Search and Bayesian Optimization for hyperparameter tuning, while also introducing exchange rate features, namely the exchange rate of the rupiah against the dollar, as an additional variable. The results of the research show that the exchange rate has a positive effect on both models, with the average test results from the GRU method with the exchange rate having an R2 Score of 0.958113. GRU without the exchange rate has an R2 Score of 0.937033, while the LSTM method with the exchange rate has an R2 Score of 0.925400. LSTM without the exchange rate has an R2 Score of 0.921825.