A Hybrid LSTM-Transformer Approach for Financial Markets: Forecasting Stock Price Time Series

Abstract

Although the most used Long Short-Term Memory (LSTM) networks are now well-suited for short-term data changes, they are not enough to grasp the overall data in financial forecasting. This paper proposes a hybrid approach that combines LSTM and Transformer architectures to leverage both temporal dynamics and global dependencies in financial time series. Specifically, the model employs LSTM layers to extract sequential and local temporal features, while the Transformer’s self-attention mechanism captures long-range correlations and the global structure of the data. Experimental results demonstrate that this combined framework achieves higher prediction accuracy and greater robustness compared to traditional models. This suggests that this paper model can better handle the complex, nonlinear, and highly volatile nature of financial data. Overall, this hybrid model provides a more reasonable and reliable basis for financial forecasting, offering valuable insights for investors, analysts, and policymakers seeking to make data-driven decisions in an ever-changing market environment.