Accurate time series forecasting is crucial for a variety of real-world applications. Despite the success of Transformer-based models, their accuracy remains limited in certain scenarios. To address this issue, we propose the MambaDBF model, which combines the strengths of the efficient state-space Mamba architecture with a dual-branch framework. The model processes the input time series through two distinct branches: the variate branch and the patch branch. Each branch integrates the proposed MambaFFN Block, which replaces the convolution operation in the original Mamba block with Feed Forward Networks (FFN). This modification effectively mitigates overfitting and significantly enhances MambaDBF’s ability to capture complex temporal dependencies. Additionally, we propose the Exponentially Weighted Signal Decay Loss (EWSDL), a loss function that applies an exponential decay factor to the error weights, making the loss function more sensitive to near-future errors. Extensive experiments on public datasets demonstrate that MambaDBF outperforms existing models across various datasets. Our code will be available at: https://github.com/mieya/MambaDBF.