Comparative Analysis of Chronos-T5 Foundation Model and Mamba State Space Architecture for Urban Traffic Forecasting

Sammanfattning

Background: Urban traffic forecasting remains a critical yet complex challenge within Intelligent Transportation Systems (ITS). While large-scale time-series foundation models offer powerful generalization, their computational overhead often conflicts with the strict latency budgets required for high-frequency, real-time urban deployment. Objectives: This study aims to mathematically quantify the operational trade-offs between deterministic predictive accuracy, probabilistic distributional fidelity, and real-time computational inference latency when comparing generalized foundation models (Chronos-T5) against locally trained state-space architectures (Mamba). Methods: To ensure a rigorous comparative baseline, the Chronos-T5 foundation model was evaluated zero-shot utilizing strictly historical, univariate traffic speed from the METR-LA dataset. Conversely, the Mamba architecture was locally trained and augmented with synchronized weather covariates (precipitation and wind speed) to isolate environmental impacts. Furthermore, Mamba was implemented utilizing native, hardware-accelerated state-space kernels to accurately benchmark its real-time inference efficiency against the transformer baseline. Results: The benchmark revealed a strict operational trade-off. Chronos-T5 achieved superior point-prediction accuracy (MAE: 1.58 mph) but incurred a pro-hibitive average inference latency of 1.35 seconds per individual 1-hour forecasting window. In contrast, the custom-trained native Mamba architecture processed equivalent prediction windows in just 7.16 milliseconds. While Mamba’s deterministic point-accuracy was lower (MAE: 4.57 mph), it achieved a vastly superior probabilistic fit, recording a KL Divergence of 1.07 bits compared to Chronos-T5’s 16.45 bits. Furthermore, the ablation study confirmed that the integration of weather covariates reduced Mamba’s deterministic error by 10.57%. Conclusions: Foundation models dominate deterministic point accuracy, but their computational burden restricts high-frequency, real-time deployment. Conversely, the native Mamba architecture remains superior at capturing probabilistic variance and structural distribution at ultra-low latencies, proving it is a highly scalable and computationally efficient paradigm for ITS.