@mastersthesis{thol3239,
           month = {December},
           title = {A comparative study on aviation arrival delay prediction using machine learning methods},
          school = {Sunway University},
          author = {Pui Ting Chew},
            year = {2023},
        keywords = {flight arrival delay prediction; logistic regression; random forest; artificial neural network; model comparison},
             url = {http://eprints.sunway.edu.my/3239/},
        abstract = {Flight delays have given rise to a great deal of redundant costs. According to the Federal Aviation Administration (FAA), it is estimated that the total cost of domestic flight delays in the United States (US) has increased from USD24 billion in 2016 to USD33 billion in 2019, a 39\% increase over the years. These costs are a composition of costs to passengers and airlines, indirect costs of delay and cost of passenger decisions to avoid future air travel, in other words, lost demand. The top three delay causes are identified as aircraft arriving late, National Aviation System (NAS) delay and air carrier delay. This research aims to identify the most important features for flight delay prediction, build supervised machine learning algorithms (i.e., logistic regression (LR), random forest (RF) and artificial neural network (ANN)) for predicting flight arrival delay and compare the performances of the methods. Dataset from 2016 to 2020 with 35 variables for Southwest Airlines Co. carrier are sourced from the Bureau of Transportation Statistics (BTS) to be trained and validated as Southwest Airlines Co. holds the biggest share of number of flights as compared to other airlines across the years. The feature selection process is carried out via literature review approach and stepwise regression approach. Both approaches recommend departure delay as the most important feature for model building. From stepwise regression approach, taxi in and taxi out are important to be included as features for flight arrival delay prediction. Considering the afore-mentioned model building methods for both approaches, ANN model using stepwise regression approach performs the best for flight arrival delay prediction based on F1 score. The F1 score for ANN train set managed to achieve 0.9 and maintain at 0.81 for its test set despite having outlier data in 2020. Furthermore, 5-fold cross validation on ANN shows overall accuracy and misclassification rate of 0.952601 and 0.047399 respectively, with insignificant differences among the folds. Additionally, the research extends its modelling efforts to assess the viability of constructing a predictive model that can be effectively applied across multiple years and concludes that the 2016 ANN model using four common significant variables from literature review approach and stepwise regression approach can produce F1 scores that exceed 0.81 when used to predict test set data from years 2017 to 2019. This research bridges the gaps from other literatures by comparing ANN with other methods such as LR and RF, considering multiple years and more features for model building. Hence, the best performing method in this research can be used to predict flight arrival delays to reduce the above-mentioned flight delay costs and aid the classification of high or low risks of the insured to be used as input for future work to calculate insurance premium based on the sum of mean and standard deviation of losses.}
}