Flight Delay Prediction

Two questions, not one

The project treats delay as both a classification and a regression problem: will this flight be delayed, and if so by how much. Those need different metrics — ROC-AUC and F1 for the first, mean absolute error for the second — and the final models report 0.73 ROC-AUC, 0.61 F1 and roughly 30 minutes MAE.

Why interpretability was the deliverable

A delay prediction that nobody can act on is a curiosity. The useful output is knowing which delay causes are controllable by the operator and which are not — weather is not actionable, turnaround scheduling is.

SHAP values were used to attribute each prediction to its contributing features, which turns the model from a black-box score into an argument about where operational effort would actually pay off. The recommendations that came out of the project rest on that attribution rather than on the raw predictive scores.