Prediction of airport on-time performance

We investigate the factors contributing to departure and arrival delays at a major international airport and develop predictive models to estimate both the likelihood and duration of delays. Using logistic regression, random forest, and gradient boosting methods, we identify key predictors of flight punctuality, including historical delay rates of flight numbers and airlines, weather conditions, runway traffic, walk time from security to gate, and overall airport congestion. Our models achieve strong inference and predictive performance in both classification and regression tasks, demonstrating the potential for targeted operational interventions to improve on-time performance and providing actionable insights for airport management and airline operations.

💡 Research Summary

This paper investigates the determinants of departure and arrival delays at a major international airport and develops predictive models to estimate both the likelihood of a delay and its duration. Using a comprehensive dataset covering all flights between 2019 and 2025, the authors compile over 850,000 valid observations after cleaning out cancelled flights, operational outages, and the anomalous period of the COVID‑19 pandemic. Each record contains flight number, scheduled and actual times, origin and destination airports, airline, aircraft type, gate assignment (including remote gates), and the walking time from security to the gate. Weather data (temperature, wind speed, visibility, precipitation) are collected hourly, together with six‑hourly forecasts extending up to 48 hours. Additional airline‑level attributes (fleet size, number of destinations, alliance membership) and airport‑level load factors are also incorporated.

The study distinguishes two modeling objectives. First, inference models aim to uncover causal drivers of delays; therefore they use “instant” features such as actual weather and runway traffic, while deliberately excluding proxy variables like historical OTP that are correlated but not causal. Second, prediction models are designed for operational use; they incorporate forecasted weather, historical delay statistics per flight number, airline, and gate, and other variables available at the time of decision‑making. Both objectives are addressed with classification (on‑time vs delayed) and regression (delay minutes) tasks. Three algorithms are evaluated: logistic regression (baseline), Random Forest, and Gradient Boosting Machines (GBM). Hyper‑parameters are tuned via cross‑validation, and model performance is assessed with weighted‑average F1‑score for classification (to account for the imbalance where on‑time flights constitute ~80 % of the data) and with MAE/RMSE for regression.

Results show that the Random Forest classifier achieves a weighted‑average F1‑score of 0.86, substantially outperforming logistic regression. The GBM regression model attains MAE = 6.3 min and RMSE = 12.4 min, a 15‑20 % improvement over traditional statistical baselines. Permutation feature importance reveals that the most influential predictors are: (1) historical delay rates by flight number and airline, (2) forecasted precipitation type (rain, snow, freezing rain, ice pellets), (3) runway congestion measured by concurrent take‑offs/landings, (4) walking time from security to gate, and (5) forecasted wind speed. Notably, historical delay rates exhibit strong interaction effects, indicating that delays propagate through the network of flights. Weather variables retain high importance in both actual and forecasted forms, but model performance degrades sharply when forecast accuracy drops, underscoring the need for reliable meteorological inputs.

The authors discuss operational implications. By identifying high‑risk flights and airlines, airport operators can apply schedule padding, prioritize gate assignments, or adjust runway sequencing during adverse weather. Real‑time decision support tools can be built on the probabilistic outputs of the Random Forest classifier, allowing stakeholders to gauge the probability of a delay rather than a binary outcome. Moreover, the identified key drivers can be turned into performance indicators for continuous monitoring and policy evaluation.

Future work is suggested in two directions: (1) extending the modeling framework to graph‑based neural networks that capture inter‑airport delay propagation over longer horizons, and (2) integrating reinforcement‑learning approaches for dynamic scheduling and resource allocation, thereby translating predictive insights into actionable, optimal operational strategies.