LTSMamba: A Long-Term Time Series Forecasting Model Based on Mamba-2

Xiyang Peng, Ruoxin Ni, Yuncheng Jiang

Abstract: Time series forecasting is extensively used in various scenarios, such as traffic forecasting and industrial management. Numerous deep learning models have been proposed for long-term time series forecasting. In particular, Transformer-based models have shown significant efficacy as they can capture long-term temporal dependency. However, existing forecasting models primarily focus on efficiently modeling the long-term temporal dependency yet often overlook the high-dimensional inter-series correlations, which are also crucial for improving the performance of long-term time series forecasting. To fill the gap, we propose LTSMamba, a long-term time series forecasting model based on Mamba-2. Specifically, LTSMamba consists of two primary components: the Temporal Dependency Block (TDB) and the Inter-series Correlations Block (ICB). TDB leverages the Bidirectional Mamba-2 Layer to capture long-term temporal dependency, while ICB utilizes Mamba-2 to extract high-dimensional inter-series correlations with an adaptive re-ranking strategy. Adaptive re-ranking is designed to adjust the time series to obtain the continuous representations, which leverages the inherent continuity bias of Mamba-2 to enhance feature learning. Extensive experiments conducted on eight real-world benchmarks demonstrate that LTSMamba outperforms state-of-the-art methods both in forecasting performance and computational efficiency.