PERBANDINGAN MODEL DEEP LEARNING RECURRENT NEURAL NETWORK DAN LONG SHORT-TERM MEMORY DALAM PREDIKSI HARGA SAHAM (STUDI KASUS: PT TELEKOMUNIKASI INDONESIA TBK MENGGUNAKAN DATA HISTORIS TAHUN 2014–2024)

Time series data are widely used in analysis and forecasting by utilizing historical patterns to predict future values. Various classical linear models, such as ARIMA, are commonly applied due to their strong mathematical foundation but are limited in capturing complex, nonlinear, non-stationary, and volatile patterns, such as stock price data. Therefore, deep learning approaches, particularly Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM), are used because they can model nonlinear relationships in sequential data. This study aims to explore the steps of RNN and LSTM models, implement both models on PT Telkom Indonesia Tbk stock price data for the period 2014–2024, and compare their performance in predicting stock prices. Both models were developed with the same architecture, consisting of two hidden layers, 64 neurons, tanh activation function, and batch size of 64. Performance evaluation was conducted using MSE, RMSE, and MAE. The results show that LSTM outperforms RNN with MSE 67.6683, RMSE 55.3865, MAE 41.5585, and training time of 497.18 seconds, whereas RNN has MSE 9810.32, RMSE 99.05, MAE 79.67, and training time of 195.24 seconds. Although LSTM requires a longer training time, it is more effective in capturing stock price movement patterns.