<![CDATA[The Indonesian capital market exhibits a high level of volatility, requiring stock price prediction models that are accurate and adaptive. Conventional forecasting models have limitations in capturing nonlinear patterns and multivariate relationships within stock time series data. This study compares the performance of Long Short-Term Memory (LSTM) and Temporal Fusion Transformer (TFT) models in predicting stock prices of LQ45 index companies, namely BBRI, TLKM, and ADRO, with a 7-day forecasting horizon. Both models were trained using 12 combined technical and fundamental features with a data split ratio of 70:15:15. Model evaluation was conducted using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). The results indicate that the TFT model achieved better performance on BBRI with a MAPE of 0.74% and TLKM with a MAPE of 0.91%, while also demonstrating faster training convergence compared to LSTM. In contrast, the LSTM model outperformed TFT on ADRO with a MAPE of 2.71%, which exhibited a relatively consistent trend pattern. Overall, TFT proved to be more effective for stocks with complex multivariate dynamics, whereas LSTM remained competitive for stocks with more stable trend patterns. The selection of prediction models should therefore consider the volatility characteristics and movement patterns of each stock issuer]]>
Copyrights © 2026
