Year: 2025
Type: conference-paper
Abstract: Precise forecasting of temporal sequences holds paramount importance in sectors like financial markets, where both overarching patterns and specific fluctuations guide strategic choices. Contemporary prediction frameworks regularly struggle to strike an optimal balance between computational cost and forecasting precision, especially when handling extended and intricate data streams. Our research introduces T-Mamba, a composite framework that combines Mamba's structured state-space sequence modeling capabilities with attention-driven mechanisms from Transformer architectures to overcome these limitations. The T-Mamba design utilizes a hierarchical scaling processor that separates chronological information into broad macroscopic trends and detailed microscopic variations, which are then processed by Mamba modules and a Localized Window Transformer respectively. An innovative routing mechanism adaptively synthesizes these distinct outputs, facilitating resource-efficient yet accurate predictions. Performance assessment on NASDAQ Historical Prices information reveals that T-Mamba substantially exceeds competing methodologies including CNN, BiLSTM, LSTM, XGBoost, and conventional Transformer implementations across evaluation criteria such as MSE, RMSE, MAE, and R^2. Component analysis confirms the critical contribution of each architectural element, highlighting the framework's structural integrity. By delivering enhanced predictive capabilities while maintaining linear complexity, T-Mamba establishes a new benchmark for temporal sequence forecasting that proves particularly valuable for extensive implementations across multiple disciplines.
Authors: Junyou Chen, Beichen Fan, Qinghao Zhang
Institutions: Zhejiang Wanli University, Wenzhou-Kean University
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W4413845243