<![CDATA[This study develops and evaluates a hybrid ICEEMDAN–SVD–LSTM model for forecasting the daily closing prices of BBRI, UNVR, and UNTR over an observation period of 15 years (September 2010–September 2025). The novelty lies in applying ICEEMDAN–SVD–LSTM to stock forecasting on the still underexplored Indonesia Stock Exchange, and in systematically comparing decomposition- and denoising-based models within one uniform framework across three sectors with different volatility profiles. ICEEMDAN was chosen because it yields a cleaner decomposition, with fewer spurious modes and less residual noise than EMD, EEMD, and CEEMDAN, while Hankel-matrix-based SVD denoising increases the signal-to-noise ratio of each component before the LSTM forecasting stage. The model's performance was compared with seven benchmarks: LSTM, EMD–LSTM, EEMD–LSTM, CEEMDAN–LSTM, ICEEMDAN–LSTM, EEMD–SVD–LSTM, and CEEMDAN–SVD–LSTM. In terms of RMSE, MAE, and MAPE, ICEEMDAN–SVD–LSTM produced the lowest errors for all three stocks, with RMSE values of 16.4605 (BBRI), 24.3086 (UNVR), and 70.8494 (UNTR), and lower MAE and MAPE than every benchmark. A Diebold–Mariano test confirmed that this advantage is statistically significant against all benchmark models. These results indicate that the proposed model is effective for short-term stock price forecasting and can serve as a decision-support tool for investors and analysts, while acknowledging market risk.]]>
Copyrights © 2026
