Predicting Airline Departure Delays — Technical Screen Prompt

Context

You have 15 minutes to review a slide deck on predicting airline departure delays, followed by a 20-minute presentation to the Director of Operations. The current analysis:

• Uses linear regression with the date encoded as an integer.
• Includes plots that mix weekdays and months without controlling for seasonality.

Assume you have historical flight-level data (schedules, realized times), route metadata, and access to weather data. The goal is to improve predictive accuracy and translate insights into operational actions.

Tasks

1. Critique the current feature engineering, including:
   • How to encode time (date as categorical or cyclical features).
   • Weather joins and availability.
   • Route- and airport-level effects.
   • Leakage risks (e.g., using wheels-off time).
2. Propose a modeling approach, covering:
   • Whether to predict continuous delay (regression) or classify delay > 15 minutes (or both).
   • Appropriate metrics (e.g., RMSE/MAE for regression; AUC-PR/cost-sensitive metrics for classification).
   • Cross-validation that respects time order.
3. Identify two charts to replace from the current deck, explain the replacements (how and why), and provide a concise narrative for a non-technical audience that links insights to actions (e.g., crew buffers by route and time-of-day).
4. Deliver a concrete recommendation to reduce compensation payouts by 5% without increasing cancellations, and define an experiment or KPI plan to verify impact within four weeks.