DuMambaformer: Hybrid state-space model with dual-branch architecture for predicting soccer match outcomes

Predicting soccer match outcomes is a fundamental task in sports analytics, enabling evidence-driven decision-making for coaches, scouts, and analysts in match preparation and performance management. The task is challenging due to heterogeneous pre-match signals, strong nonlinearity in team interactions, and long-range dependencies across season-level histories, where decisive events are sparse and easily diluted. Existing methods span statistical and rating approaches (e.g., Elo- and Poisson-based models), classical machine learning pipelines (e.g., gradient-boosted decision trees), and deep sequence architectures such as convolutional neural networks and attention-based Transformers. However, traditional models often rely on limited assumptions and struggle to capture the complex temporal evolution of systems with heterogeneous features. Attention-centric models can model interactions but incur O ( N 2 ) computational and memory costs with respect to history length, hindering low-latency deployment for long horizons. Efficient state-space models reduce complexity but may underperform when precise content-based retrieval and cross-match alignment are needed. To address these limitations, we propose DuMambaformer, a hybrid state-space and attention model with a dual-branch architecture that efficiently captures long-context temporal dynamics while preserving decisive cues. DuMambaformer integrates a parameter-shared dual-branch Mambaformer temporal pathway with gated fusion, a dilated feature-axis convolutional pathway for multi-scale cross-feature interactions, and an alignment-based fusion module followed by a lightweight classification head. Experiments on two public real-world datasets, Football-Data and FiveThirtyEight Soccer SPI, demonstrate the effectiveness and robustness of DuMambaformer. Ablation studies further verify the contributions of temporal modeling, spatial feature extraction, alignment fusion, and weighted focal loss, while efficiency results indicate competitive inference latency under fixed history windows.