TechPrufer
Sign inGet started

Blogs›Uber›Real-time Traffic Forecasting

Real-time Traffic Forecasting

Real-time Traffic Forecasting

1
posts
2026

Uber has significantly enhanced its real-time traffic forecasting system, initially built over a decade ago. The latest iteration, DeepETT (Deep Estimated Travel Time), leverages deep learning, specifically a graph-aware transformer model, to improve arrival time accuracy, especially for longer trips. This multi-year effort focused on addressing challenges in rapidly changing traffic conditions, sparse road networks, and the compounding effects of segment-level inaccuracies on trip-level ETAs. The system now serves upwards of 2 million real-time forecasts per second, improving long-trip arrival time accuracy by 6% and boosting forecast variance explained by 19%. Key engineering decisions included decoupling training from the routing loop and using fixed-size inputs from pre-aggregated observations to maintain predictable inference and scale. The system defines contracts for segment-level and trip-level forecasts, decomposing metrics into resolution and calibration to manage the paradox of segment MSE improvements not always directly translating to trip MAE improvements. The model design utilizes local receptive fields with spatiotemporal views and context features to forecast segment-level ETTs.

2026

Scaling Real-Time Traffic Forecasting with a Graph-Aware Transformer

5/19/2026

This post details the engineering decisions behind DeepETT, Uber's new real-time traffic forecasting system powered by a graph-aware transformer model. It highlights the system's ability to improve long-trip arrival time accuracy and forecast variance explained, serving millions of forecasts per second. Key technical contributions include the decision to decouple training from the routing loop and use fixed-size inputs from pre-aggregated observations for predictable inference and scalability. The post also discusses the system's contracts for segment-level and trip-level forecasts, the decomposition of metrics into resolution and calibration, and the model's design utilizing local receptive fields with spatiotemporal views and context features.