AeroFlux: Stage 1 — Predictive ML Model
A production-scale, propagation-aware ML pipeline trained on 61M+ U.S. domestic flight records (2015–2024) from the Bureau of Transportation Statistics, fused with NOAA hourly weather and airport metadata under strict temporal alignment (no look-ahead bias).
Core technical components:
- Data fusion pipeline — BTS flight records, NOAA weather observations, and airport metadata joined via UTC-normalized as-of joins, with engineered tail-rotation, congestion, and propagation features
- XGBoost production model — gradient-boosted trees (300 trees, depth 6, η = 0.05) trained on 61M flights in ~17 minutes using the histogram method, with dual classification/regression heads
- LSTM model — dual-output (classification + regression) over 3-timestep aircraft rotation sequences; built but deemed too memory-intensive for full multi-year deployment
- Time-aware rolling-origin cross-validation — 5 expanding-window folds (2015 → 2023) validating on subsequent years, with 2025 as a true unseen holdout
- Pre-departure constraint — no departure-delay leakage; uses only features available before the aircraft leaves the gate
Headline Results
On the 2025 holdout (6.86M unseen flights) after tuning:
| Metric | Value |
|---|---|
| AUC | 0.840 |
| F₁ | 0.554 |
| MAE | 20.3 min |
| Precision | 79% |
When the model issues a delay alert, it is right 4 out of 5 times — operationally useful for rebooking triage. Across 5 rolling-origin folds (2019–2023 validation years), mean AUC was 0.836 and mean MAE was 18.1 minutes.
The Dominant Signal
51% of total model importance comes from one feature: prev1_arr_del15 — was the same aircraft late on its previous leg? This single finding validates the central thesis: propagation is the signal. Upstream aircraft state dominates downstream delay outcomes. Most flight-delay models ask what is true about this flight? AeroFlux asks what state is the aircraft carrying into this flight?
Try It & Read More
Open the OR568 project site on GitHub Pages — full methodology, ML workflow notebooks, and paper
OR568 Project repo on GitHub — data pipeline, ML pipeline, and reproducible environment
AeroFlux repo on GitHub — the inference system and web app