<![CDATA[This study compares the performance of three machine learning algorithms, Long Short-Term Memory (LSTM), Random Forest, and Extreme Gradient Boosting (XGBoost), in predicting the next-day closing price of Indonesia's Composite Stock Price Index (IHSG) as a baseline before further feature engineering is applied in a broader ongoing study. Daily historical price data (Open, High, Low, Close, Volume) covering January 2015 to early 2026 were collected from Yahoo Finance. Two feature representations were compared: the raw 5-dimensional OHLCV attributes, and a 64-dimensional temporal representation extracted from the same OHLCV data using a Bidirectional LSTM (Bi-LSTM) encoder. Each representation was evaluated using LSTM, Random Forest, and XGBoost, with performance measured by Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) on a chronological 80:20 train-test split. The results show that the raw OHLCV representation combined with LSTM achieved the best performance (RMSE = 196.11, MAE = 164.14, MAPE = 2.18%), outperforming Random Forest and XGBoost on the same representation (MAPE 3.78% and 3.80%). Encoding the OHLCV data into a 64-dimensional Bi-LSTM representation without any external signal reduced accuracy across all three algorithms, most notably for LSTM (MAPE rising to 7.13%), indicating that unsupervised temporal encoding discards useful absolute price information when no additional predictive feature is introduced. These findings establish a validated baseline configuration and evaluation pipeline for IHSG closing-price prediction, intended as the foundation for a subsequent study that integrates external textual features into the same experimental framework. ]]>
Copyrights © 2026
