Application of LSTM-Based Deep Learning for Stock Return Prediction of DCII

Abstract

Stock return prediction is one of the areas that has received great attention in modern finance because it can help investors make more informed decisions and reduce the risk of market uncertainty. This study applies a deep learning approach based on Long Short-Term Memory (LSTM) to predict the return of DCII (PT DCI Indonesia Tbk) shares as a representation of highly volatile stocks on the Indonesia Stock Exchange. The purpose of this study is to evaluate the performance of twelve LSTM variants—including LSTM-Base, LSTM-Wide, LSTM-Stack2, LSTM-Stack3, LSTM with Dropout, BiLSTM, BiLSTM with Attention, and LSTM with Attention Mechanism—by comparing their performance on daily (H=1) and weekly (H=7) prediction horizons using historical data from id.investing.com. The initial data undergo preprocessing involving local format cleaning, calculation of technical indicators (MA, EMA, MACD, RSI, ATR, Bollinger Bands, etc.), MinMax normalization, and sequencing (windowing) with 30, 60, and 120-day lookbacks. The training process uses a uniform configuration with Adam optimization and early stopping to prevent overfitting, while the evaluation employs MAE, RMSE, MAPE, and R² metrics. The results show that LSTM-Stack3 (LB=60, H=1) provides the best performance with MAE = 0.020, RMSE = 0.031, MAPE = 5.0%, and R² = 0.91, followed by LSTM-Stack2-DO as the second-best configuration. Meanwhile, the LSTM-LB120-H7—the only model evaluated with a seven-day horizon—achieves the lowest performance due to higher long-term uncertainty. These findings confirm that stacked LSTM architectures are more effective for short-term return forecasting, whereas longer horizons require hybrid or enhanced approaches for stable performance..

GS Cited Analysis