Peningkatan Akurasi Nilai Harga Saham Menggunakan Metode Long Short-Term Memory (LSTM) pada PT Unilever Tbk

Abstract

The rapid development of technology has an impact on the economy of society, one of which is investing in stocks. Stocks are a proof of an individual's ownership of an asset in a company. However, stock prices have a very high level of fluctuation, so an accurate method is needed to help predict stock prices. LSTM and GRU were chosen due to their intrinsic ability to handle long-term and short-term issues in time series data. LSTM has a complex memory structure that allows decision-making based on long-term and short-term information. Meanwhile, GRU has a simpler structure with a focus on gate mechanisms to control the flow of information, resulting in a lighter and faster model. Therefore, this study will compare two RNN methods, namely Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU), in predicting stock prices using the stock price data of PT. Unilever (UNVR) with evaluation metrics MAPE and RMSE. The combination of parameters used to evaluate the MAPE and RMSE values in this study includes learning rate, timesteps, batch size, and epoch. The results of this study indicate that the GRU method is more accurate compared to the LSTM method. This is evidenced by the evaluation results of the LSTM method with the lowest MAPE value of 2.42% and the lowest RMSE value of 0.01807, while the evaluation results of the GRU method with the lowest MAPE value of 2.14% and the lowest RMSE value of 0.01775. The combination of parameters used in this study also has an impact on the final MAPE and RMSE results, especially with the use of learning rates of 0.001 and 0.0001. Thus, it can be concluded in this study that the GRU method is more accurate and effective compared to the LSTM method in predicting stock prices.