Will a US flight arrive 15+ minutes late? A tuned XGBoost model, predicting at gate pushback, replayed on
June 2023: 334,950 flights it never saw during training or tuning.
90.8%Accuracy (June test)
0.943ROC-AUC
0.885Macro-F1
70.7%Baseline: always "on time"
Loading…
Sched
Flight
To
Dep delay
P(delay)
Predicted
Actual
Select a flight
Click a row to see how the model's delay risk changes with departure delay and weather.
Real weatherThunderstorms at both airportsDecision threshold
Each curve changes only the departure delay (and, for the orange curve, the weather). Every
other input stays at this flight's real values.
How this demo works. The model is a tuned XGBoost classifier trained on Jan–Apr 2023
BTS flights joined with hourly airport weather, using 73 engineered features (time, route & congestion,
aircraft rotation, airport state, weather, departure delay). It was tuned on May and evaluated once on June.
This page is static: every probability shown, including the what-if curves, was precomputed by the real
model for each June flight. It replays held-out 2023 data; it is not a live forecasting service.
Code, notebooks and full write-up:
github.com/prakash-nitc/FlightDelayPrediction.