Forecasting Multivariabel Harga Saham Menggunakan Model LSTM Multi-output

Abstract

Stock price forecasting is a crucial issue in the investment world because stock price movements are volatile and influenced by various economic factors and market sentiment. Therefore, a forecasting method capable of capturing nonlinear patterns and long-term dependencies in time series data is necessary. This study aims to develop and compare the performance of single-output Long Short-Term Memory (LSTM) and multi-output LSTM models in stock price forecasting, as well as to analyze the differences between single-step and multi-step forecasting approaches. The data used in this study is historical stock price data for PT. Gudang Garam Tbk (GGRM.JK) from January 2020 to November 2025, obtained from Yahoo Finance, with Open, High, Low, and Close variables. This study built three LSTM models: two single-output models to predict the Open and Close prices separately, and one multi-output model to predict the Open and Close prices simultaneously with High and Low input variables. All data underwent preprocessing stages including data cleaning, normalization, and time series data generation with a 90-day window. Model evaluation was conducted using Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE), and two forecasting strategies were applied: single-step and multi-step forecasting. The results showed that the single-output LSTM model performed best in forecasting the Open price with a MAPE of 1.11%, while the multi-output LSTM model performed better in forecasting the Close price with a MAPE of 2.98% due to its ability to utilize the interrelationships between stock price variables. The comparison of forecasting strategies showed that single-step forecasting was more stable and accurate for short time horizons, while multi-step forecasting had a higher error rate due to error accumulation, but better represented real-world forecasting conditions. Reliability analysis showed that the multi-output LSTM model remained reliable for up to 10 days with a MAPE of 0.99%. Implementation on the following month's data shows that the model remains suitable for use as a stock price forecasting tool.