Font Size: 

Short-term road traffic forecasting is a key function of intelligent transportation systems and connected and automated mobility, supporting congestion management, route guidance, V2X-enabled traffic control, and cooperative automated driving services. However, roadside sensing networks often contain missing observations caused by detector malfunction, communication dropout, maintenance activity, and power interruptions, conditions that are rarely addressed in standard model comparisons. This paper presents a systematic evaluation protocol for missing-data robustness across representative deep learning architectures for road traffic forecasting, covering recurrent, graph-convolutional, diffusion-based, attention-based, and hybrid recurrent–graph models, specifically GRU, STGCN, DCRNN, ASTGCN, Graph WaveNet, and AGCRN. Forecasting accuracy is measured at 15-, 30-, and 60-minute horizons using MAE, RMSE, and MAPE, computed with standard masking of invalid sensor values. Beyond standard complete-data evaluation, two controlled missing-data regimes are introduced to simulate random point missingness and sensor-block failures. In both cases, missingness is applied only to model inputs, while ground-truth prediction targets are drawn from the original complete data. The evaluation design compares absolute forecasting error with relative degradation from complete-data performance, providing a structured basis for assessing architectural robustness under realistic sensing conditions. The study is intended to support principled model selection for intelligent transportation and connected vehicle deployments where resilience to imperfect sensor streams is critical for dependable operation.