Uber Deploys DeepETT, a Deep Learning Traffic Forecasting System Serving 2M+ Forecasts Per Second and Driving $100M Annual Revenue Gains
Key Takeaways
- ▸DeepETT improves arrival time accuracy by 6% and increases forecast variance explained by 19%, with an estimated $100M annual revenue impact
- ▸The system processes 2+ million real-time traffic forecasts per second, making it one of Uber's highest-throughput deep learning deployments
- ▸Graph-aware transformers enable sophisticated understanding of road networks and traffic interdependencies across 100+ million segments globally
Summary
Uber has announced DeepETT (Deep Estimated Travel Time), a deep learning-based real-time traffic forecasting system representing a multi-year rebuild of its foundational traffic prediction infrastructure. The new system improves long-trip arrival time accuracy by 6%, increases forecast variance explained by 19%, and generates an estimated $100 million in incremental annual revenue. Operating at one of Uber's highest-throughput deep learning deployments, DeepETT serves over 2 million real-time forecasts per second globally.
The system leverages graph-aware transformers to process tens of billions of daily GPS location updates, mapping them onto a global road network of approximately 100 million segments. DeepETT was architected to address three critical gaps in Uber's prior forecasting approach: rapidly adapting to changing conditions (incidents, weather, events), improving generalization across sparse roads outside dense urban areas, and enhancing accuracy for longer trips where segment-level improvements compound significantly.
Key to the system's success was a pragmatic engineering decision to maintain traffic forecasting as a separate module from the routing optimization engine, prioritizing operational reliability and global deployability over theoretical elegance. This architectural choice, combined with advanced deep learning techniques and careful feature engineering, enabled DeepETT to scale reliably across Uber's worldwide operations while adapting quickly to diverse traffic patterns and geographies.
- Uber prioritized engineering pragmatism and deployability over theoretical optimization, keeping traffic forecasting separate from routing for operational reliability
Editorial Opinion
DeepETT exemplifies how sustained investment in foundational ML infrastructure can deliver massive business value at planetary scale. Uber's decision to trade theoretical elegance for engineering pragmatism—keeping traffic forecasting separate from routing—is a valuable lesson for enterprise AI deployments. The achievement of serving 2M+ forecasts per second while driving $100M in revenue impact demonstrates that real-world infrastructure problems remain one of the most compelling applications of deep learning.
