We trained an XGBoost model on U.S. domestic flight records from the Bureau of Transportation Statistics — 601K+ flights across 14 carriers — accurate within ±30 minutes 86.5% of the time.
The project
SkyPredict isn't another dashboard template. It's a research project that combines 601,570 BTS flight records, 257K hourly weather observations, and a tuned XGBoost model to estimate whether a given domestic flight will depart late.
The model considers 45 engineered features — including cyclical time encodings, airport congestion proxies, carrier delay rates, and holiday indicators — then outputs a delay probability and an estimated delay in minutes.
Capabilities
Pick an airline, route, and departure time. The model classifies delay risk and — if a delay is likely — estimates how many minutes to expect. Uses a dual-model architecture: classifier → regressor.
See which airlines, routes, and hours have the worst delay rates. Interactive charts built from the same BTS dataset.
Enter a flight number to pull real-time gate, departure, and arrival data from the AviationStack API.
Under the hood
Raw inputs (airline, airports, date, time) are transformed into 45 features: cyclical sin/cos of hour and month, one-hot carrier codes, historical route delay averages, airport congestion windows, holiday proximity, and more.
An XGBoost classifier (75 Optuna trials, temporal Oct 1-25 / Oct 26-31 split) predicts whether the flight will be delayed (>15 min). Threshold tuned for best F1.
If the classifier predicts delay, a second XGBoost regressor estimates the expected delay duration in minutes. Both models share the same feature pipeline.