Personal Project · Machine Learning
Flight Delay Prediction
A machine learning study of 251,780 Southwest Airlines departures from LAX (2017–2024, Bureau of Transportation Statistics). I first tried predicting delay minutes with regression, then reframed the problem as classifying flights that leave at least 15 minutes late — documenting every experiment along the way.
- Cleaned the raw records with pandas and engineered calendar, scheduled-departure-time, and cyclical features, keeping outcome information out of the inputs to prevent leakage.
- Used chronological train / validation / test splits so models are always evaluated on future flights, the way they'd be used in practice.
- Compared Linear Regression and Random Forest regressors against a median baseline on MAE, then trained Logistic Regression and Random Forest classifiers in scikit-learn pipelines.
- Tuned the decision threshold on validation F1 to balance precision and recall, and showed why accuracy alone is misleading when most flights are on time.
- Generated figures and Markdown reports from saved results, and saved the final pipeline and threshold with joblib for new predictions.