Flight delay forecasting for the 50 busiest U.S. routes using XGBoost, LightGBM, LSTM, and TCN. Trained on 6 years of BTS flight records and Open-Meteo weather data (2019 - June 2025).
Live dashboard: tisyasharma.github.io/flight-delay-forecasting
| Model | MAE | Hit Rate |
|---|---|---|
| XGBoost | 11.25 min | 77.6% |
| LightGBM | 11.25 min | 77.7% |
| LSTM | 12.69 min | 74.42% |
| TCN | 12.79 min | 72.65% |
| Naive Baseline | 15.09 min | 67.7% |
| Moving Average (28-day) | 13.53 min | 70.0% |
Gradient boosting reduced MAE by 25% compared to the naive baseline (11.25 min vs 15.09 min) and outperformed deep learning on every route tested. Metrics averaged across four walk-forward evaluation windows (2023-2024). Removing weather features increased error by 10.3%, the largest impact of any feature group.
- Flights: BTS On-Time Performance (Jan 2019 - Jun 2025)
- Weather: Open-Meteo ERA5 reanalysis, hourly data aggregated into daily operating-hour metrics
Notebooks 01-08 walk through EDA, feature engineering, model training, and error analysis. The dashboard includes interactive breakdowns by route, season, and model.