Machine learning approaches to airfare prediction : a comparative study with hybrid models

Abstract

Due to the dynamic pricing strategy and market conditions, the price of airfare fluctuates largely and is difficult to predict based on common knowledge by customers, making it difficult for travelers to find the best time to order a flight ticket a challenge. This thesis explores machine learning and deep learning approaches on airfare prediction to achieve better performance. A real-world dataset consisting of Indian domestic flight data integrated with extra external features such as fuel price and holiday flag in both tabular and time series versions were utilized in the thesis. Machine learning and deep learning models including Random Forest, XGBoost, GRU, LSTM and BiLSTM were evaluated based on MAE, RMSE and R^2. A hybrid GRU-XGBoost model and a meta-learning pipeline outperformed others, with the best R² scores of 0.9469 (for tabular data) and 0.6938 (for time series data). The results showed the effectiveness of applying hybrid-architecture and structured features in airfare prediction.