MAS‐Mamba: A Compact Missing‐Aware Selective State‐Space Model With Calibrated Uncertainty for Short‐Term Traffic Forecasting Under Sensor Outages

ABSTRACT

Short‐term urban traffic forecasting in operational deployments is frequently affected by sensor faults, communication outages and irregular sampling, which produce multivariate time series with structured missingness. Existing graph‐ and Transformer‐based forecasters often rely on cleaned inputs or upstream imputation, and recent selective state‐space models (SSMs), including Mamba, do not explicitly use observation masks inside the latent‐state transition. This study develops MAS‐Mamba, a compact missing‐aware selective SSM in which the input‐dependent discretisation step is conditioned on the observation mask and the elapsed time since the last valid reading, so that the latent state is preserved or damped during sensor outages instead of absorbing imputed values. Three SSM branches at 5‐, 15‐, and 45‐min resolutions are fused by a cross‐scale gate, and a heteroscedastic split‐conformal head produces prediction intervals with finite‐sample marginal coverage under exchangeability. We additionally derive a conditional latent‐state stability bound that motivates the mask‐aware design; we emphasise that it bounds latent‐state deviation under stated assumptions and is not a guarantee on prediction error. The model is evaluated on the full 207‐sensor METR‐LA and the 325‐sensor PEMS‐BAY benchmarks under controlled MCAR and NMAR missingness, against compact baselines and strong spatio‐temporal graph models (DCRNN, Graph WaveNet), with a controlled benchmark (CMTB) used only for mechanism analysis. MAS‐Mamba consistently improves over a vanilla selective SSM and is competitive with GRU‐Lite at a much smaller parameter budget; a lightweight road‐graph variant (MAS‐Mamba + Spatial) further improves accuracy. Full graph models attain the best point accuracy when a reliable road graph is available, while MAS‐Mamba offers a 40–90 smaller, missing‐aware alternative with empirical conformal coverage close to the 90% and 95% nominal levels across both datasets. The results position MAS‐Mamba as a compact, missing‐aware, parameter‐efficient building block with calibrated uncertainty, rather than as a claim of general dominance.