Flight Delay Propagation Intelligence System
Predict the cascade.
Act before the chaos.
One late aircraft is never one late flight. The same airframe flies five more legs that day, and every minute it cannot recover on the ground is a minute handed to the next departure. FDPIS scores that risk before the day begins, then traces the delay through the rotation network as it happens.
One aircraft, one morning
A 34 minute delay at Dallas, traced through the rest of N974AA's day.
- DFW → EWR09:3134 min late
- EWR → ORD13:51+42 min
- ORD → DFW17:46+13 min
- Confidence boundaryDFW → MCO21:05cascade continues
Past depth 2 the model reports that the cascade continues, without a figure. That boundary is measured, not chosen.
Validated results
Measured on held-out US domestic operations
Full ablation, model comparison and propagation accuracy on the model performance page, including the limitations we have not solved.
Architecture
Five layers, from raw schedule to a dated action
Each layer consumes the one before it. The rotation graph is what makes the difference: without it the model is barely better than the schedule itself.
- 01
Data Ingestion
Schedule, actual times and tail assignments normalised into a single leg-level record.
- 02
Graph Construction
Legs linked into rotations by aircraft, so every flight knows what flew it in.
- 03
Primary Risk
Each leg scored for delay risk. Rotation features carry most of the signal.
- 04
Propagation
Observed delay pushed downstream, absorbed by turnaround slack at each stop.
- 05
Recommendations
Threshold crossings become dated actions: crew, ground handling, aircraft, slots.
See what one delay costs you by lunchtime
Enter a delay on any aircraft in the fleet and watch it move through the rest of that airframe's day, with the actions it triggers and the deadlines they carry.