Implementation Of LSTM, GRU, And BI-LSTM Algorithms For LQ45 Index Prediction

Abstract

The Indonesian capital market, particularly the LQ45 Index and its leading sectoral stocks, exhibits high volatility and complex non-linear price patterns. This complexity renders conventional statistical methods insufficient for accurate forecasting and risk mitigation. This study aims to develop and compare three Deep Learning architectures Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (Bi-LSTM) in predicting the daily close prices of the LQ45 Index and six highly liquid stocks. The research methodology adopted the CRISP-DM framework utilizing historical data from January 2020 to May 2026. To prevent data leakage and ensure scientific rigor, model training incorporated a 5-Fold Walk-Forward Validation with an expanding window and multi-seed training. Performance was evaluated and benchmarked against statistical baseline models (Naïve, ARIMA, ETS) using RMSE, MAPE, and Directional Accuracy (DA). Experimental results reveal that all Deep Learning models achieved highly accurate performance with an average MAPE below 5%. The GRU model proved to be the most optimal, recording the lowest RMSE across the majority of assets. While statistical models showed slight superiority in minimizing short-term nominal errors, Deep Learning architectures consistently outperformed in recognizing market trend directions. Finally, the best models were successfully deployed into StockAI, an interactive web application built on the Flask framework using Model-View-Template (MVT) architecture for real-time live inference.