Related Experiment Videos
EMAT: Enhanced Multi-Aspect Attention Transformer for Financial Time Series Forecasting
Yingjun Chen1, Wenfeng Shen1, Han Liu1
1School of Computer and Information Engineering, Shanghai Polytechnic University, Shanghai 201209, China.
Entropy (Basel, Switzerland)
|October 28, 2025
Summary
This study introduces the Enhanced Multi-Aspect Transformer (EMAT) for stock market prediction, improving financial time series forecasting. EMAT effectively captures complex market dynamics, outperforming existing models in accuracy.
Area of Science:
- Quantitative Finance
- Machine Learning
- Deep Learning
Background:
- Financial time series prediction is challenging due to non-stationarity and complex dependencies.
- Traditional methods struggle with multifaceted market dynamics like temporal proximity, trend, and volatility.
Purpose of the Study:
- To propose a novel deep learning architecture, the Enhanced Multi-Aspect Transformer (EMAT), for stock market prediction.
- To address limitations of existing models in capturing complex financial market characteristics.
Main Methods:
- Developed EMAT, a deep learning model with a Multi-Aspect Attention Mechanism.
- Incorporated an encoder-decoder structure with SwiGLU activation and a multi-objective loss function.
- Evaluated EMAT on multiple stock market datasets against state-of-the-art baselines.
Main Results:
- EMAT consistently outperformed various baseline models, including recurrent, hybrid, and Transformer architectures.
- Ablation studies confirmed the critical contribution of each component within the Multi-Aspect Attention Mechanism.
- Demonstrated significant improvements in predictive accuracy for financial forecasting.
Conclusions:
- EMAT is an effective and robust tool for financial forecasting.
- The model's ability to simultaneously model temporal decay, trend dynamics, and volatility regimes enhances predictive power.
- EMAT offers substantial accuracy improvements over existing financial prediction approaches.
