A Comparative Eleven-Model Benchmark forAirline Route Profitability Prediction and SeasonalCost Analysis

Abstract

Accurate airline route profitability prediction is critical for fleet allocation, scheduling, pricing, and network planning under volatile operating conditions. This paper presents a comparative machine learning framework for flight-level profit margin prediction using a public airline route profitability dataset containing 7,974 flights across 30 routes in 2024. The study integrates exploratory operational analysis with predictive modeling to investigate the impact of seasonality, route category, demand level, and aircraft assignment on airline profitability. To preserve practical planning utility, direct revenue and total cost variables are intentionally excluded from the final feature set, allowing the models to infer profitability from operational and market-related characteristics rather than accounting identities. Eleven machine learning models are systematically benchmarked, including linear regression, ensemble learning, and gradient-boosting approaches. Experimental results demonstrate that boosted tree models consistently outperform traditional linear methods. Among all compared models, CatBoost achieves the best predictive performance with test R^2=0.7042, RMSE =20.94, and MAE =15.80, followed closely by Gradient Boosting and HistGradient Boosting. Seasonal analysis further reveals that prediction errors increase from Peak to Low season, indicating that weak and volatile operating environments are more difficult to model accurately. The results confirm that operational features contain sufficient predictive signal for practical profitability forecasting and demonstrate the effectiveness of gradient-boosting methods for airline decision-support applications.