How We Predict Transit Times: Proprietary ML on 10+ Years of Network Data

The problem with generic ETA models

Every supply chain platform claims ML-powered ETAs. Most are wrappers around the same inputs: GPS coordinates, posted speed limits, and real-time traffic data from Google or HERE Maps. These models answer a driving question: how long to get from A to B right now?

Supply chain ETAs need to answer a different question: when will this specific shipment, carried by this specific carrier, on this specific lane, with this specific load type, actually arrive? A generic model does not know that Carrier X consistently adds 4 hours on the Memphis-to-Dallas lane because of a systematic dispatch delay. It does not know that Friday pickups in the Northeast have a 22% higher probability of weekend detention. It does not know that a specific shipper's Midwest facilities have a 40-minute average gate processing time. FourKites knows all of this. The network has been recording it for over a decade.

What we built

The Supply Chain Transit Prediction Model is a proprietary lane-level transit time prediction system trained exclusively on FourKites network data. Not a wrapper around a maps API. A purpose-built model that learns from the actual behavior of shipments across the network.

Training data

The model trains on 10+ years of completed shipment records. For every historical shipment, the model sees origin, destination, carrier, mode, day of week, time of day, season, load type, and the actual transit outcome. Critically, it also sees facility-level behaviors: gate processing times, appointment adherence patterns, and dwell time variations by day and shift. This training data is unique. No other platform has 11 years of cross-company shipment outcomes at this scale.

Model architecture

The baseline uses LightGBM, a gradient-boosted decision tree framework chosen for its performance on tabular data with mixed feature types. Initial phase focuses on truckload (TL) mode with single pick, single drop loads. Features include lane-level historical transit distributions, carrier-specific performance by lane, time-of-week and seasonal patterns, facility processing times (derived from TrackX yard data), and real-time signals including weather and congestion. The model outputs a predicted transit time with a confidence score.

Benchmarking and deployment

We benchmark against two baselines: current FourKites ETA logic and zero-shot foundation models (Chronos and TimesFM). The benchmark evaluates tail-end reliability, not just averages: how well does the model predict shipments most likely to be late? The model deploys first in shadow mode (predictions logged, not surfaced), then enters A/B testing with statistical significance requirements before replacing the incumbent. In supply chain, a wrong ETA can trigger a $5,000 expedite decision or a $50,000 missed production changeover. The cost of being wrong demands deployment discipline.

Why this matters

This is the first of several proprietary ML systems FourKites is building on its network data. The thesis: 11 years of cross-company supply chain outcomes constitute a training dataset that no foundation model and no competitor can replicate. Generic AI models will improve, but they will never see what happens inside Carrier X's Memphis terminal on a Friday evening. FourKites has been recording it since 2014.