FedEx improved package volume forecast accuracy by 20% with production ML

The workflow before#

FedEx moves more than 15 million shipments a day and records billions of tracking events a year. That data was already on Azure, but turning it into decisions was slow. Data science teams worked in silos with their own tools, so a promising model often stayed an experiment. Volume forecasts leaned on manual processes, and planners sized capacity from them. Customers usually learned about a delay after the package was already late, and then they called. When COVID-19 hit, demand swung by lane and service faster than the planning cycle could follow, and the gap between what the data knew and what operations did became expensive.

What we built#

Working with Oliver Wyman, we built FedEx one shared machine learning environment and took four kinds of models all the way into production. Planners get better volume forecasts, operations gets delay predictions while there's still time to act, and customers get told about a problem before they have to ask. Every model draws on the same features and ships the same way.

Forecasts by lane and service

Package volume forecasts run at the level planners actually staff to: each lane and service. Peak periods get their own attention because that's where forecast misses cost the most. We built one shared feature store so the forecasting, delay and segmentation models read the same definitions of the same events.

Delay alerts before the customer notices

A real-time model scores each shipment's risk of missing its delivery window as tracking events arrive. When the score crosses a threshold, FedEx's systems send the customer a proactive notice. We wrote predictions back into the operational systems rather than a separate dashboard, because a warning nobody acts on in time doesn't help a customer.

Scenario models for a moving network

We added COVID-19 trigger conditions and scenario models so planners could stress-test the network and move capacity as demand shifted. A planner can ask what happens if a region's volume jumps and see the effect before committing trucks and staff.

Shipping patterns and anomalies

Pattern analysis groups shippers by behavior and flags unusual volume or routing. Account and operations teams see which customers are changing how they ship, and anomalies surface for review instead of hiding in averages.

How we delivered it

The work ran 44 weeks, from data foundation through production deployment, delivered with Oliver Wyman. We built the shared environment first, then brought models through one at a time with A/B tests against the existing process, so each went live only after it beat what planners already had.

Before and after#

What's next#

With forecasting and delay models in production, the next step is acting on a predicted delay end to end. An agent could reroute or rebook a shipment at risk, notify the customer and log what it did, with a person reviewing the exceptions it can't resolve on its own.