Integrating xLSTM with Signal Decomposition for Enhanced Stock Index Prediction: A Comparative Study with LSTM on the Hang Seng Index

This paper investigates whether xLSTM, a recent extension of LSTM networks, provides advantages for financial time series forecasting. Using Hang Seng Index data, we conducted a comparative analysis of xLSTM and conventional LSTM models under multiple signal decomposition frameworks, including EEMD, CEEMD, and UPEMD. Our experimental results demonstrate that UPEMD-LSTM achieves the best performance with a MAPE of 0.52%, and UPEMD-xLSTM follows with a MAPE of 0.72%. The comparison between architectures reveals that the benefits of xLSTM are conditional on proper signal preprocessing. Consequently, architectural improvements alone are not sufficient for financial prediction. The results of this study offer evidence-based recommendations on the deployment of xLSTM architectures for financial prediction tasks and demonstrate that domain-specific preprocessing remains essential to achieve superior forecast performance.