The workflow before
FedEx moves more than 15 million shipments a day and records billions of tracking events a year. Data science teams worked in silos with different tools, and experimental models rarely reached production. Forecasts leaned on manual processes, and customers heard about delays after they happened. When COVID-19 hit, demand swung faster than the planning process could follow.
What we built
Working with Oliver Wyman, we built a machine learning environment on Azure Data Lake and Databricks over 44 weeks, from data foundation through production deployment.
We put four model families into production: package volume forecasts by lane and service, delivery time predictions, real-time delay prediction that triggers proactive customer notices, and shipping pattern analysis for segmentation and anomaly detection.
We added COVID-19 trigger conditions and scenario models so planners could test network stress and reallocate capacity as demand moved.
Before and after
| Measure | Before | After |
|---|---|---|
| Package volume forecast accuracy | Baseline | 20% better |
| Peak period forecast error | Baseline | 15% lower |
| Missed delivery windows | Baseline | 25% fewer |
| Customer calls about delays | Baseline | 25% fewer |
What's next
With forecasting and delay models in production, the next step is Level 3, Workflow Agents: acting on a predicted delay end to end, rerouting and rebooking with a person reviewing exceptions.
