An Attention-Guided Improved Decomposition-Reconstruction Model for Stock Market Prediction

The stock market plays a pivotal role in global economy. Accurate stock market prediction not only assists governments in formulating macroeconomic policies but also helps investors to make informed decisions, thereby contributing to economic and market stability. However, the inherent high volatility, nonlinearity, and nonstationarity of stock data pose significant challenges to accurate forecasting. To address these challenges, this study proposes a novel hybrid model, named Attention-guided Improved Decomposition-Reconstruction Model (AIDRM), for accurate stock market prediction. First, the model decomposes the raw stock indices using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to capture their inherent complexity. The decomposed sub-series are then concatenated and adaptively weighted through an attention mechanism, guiding the model focus on key features that affect forecasting accuracy. The concatenation of sub-series prior to prediction can mitigate the typical error accumulation problem of the traditional decomposition-reconstruction frameworks. Finally, the attention-refined features are fed into the Informer model to generate predictions. Experiments on four major indices (SSEC, SZI, STI, and S&P 500) demonstrate that the proposed model outperforms state-of-the-art deep learning benchmarks, achieving RMSE reductions of 17.6 – 32.5% across all tested stock indices. In addition, the ablation study verified the effectiveness of each module.