Prediction of airline delays

Abstract

Considering the ever-increasing air traffic, prediction of aircraft delays has become a more and more valuable tool to try to limit the negative impacts of flights running behind schedule. Better prediction of delays would lead to decreased costs, lower environment impact and better welfare for workers and travelers. Based on U.S. flight records for the year 2015, this paper aims at determining the variables that have the most impact on aircrafts delays, and also intends to analyze the various models used. Using data mining techniques such as decision trees, random forests, neural networks, regressions and support vector machines, various configurations of input and output variables are studied. The results are compared and interpreted, and the performance of the models are analyzed through indicators such as receiver operating characteristic (ROC) curves, misclassification rates, average squared errors, and lifts. Results show that the most influential variable is the time of departure, as delay propagation throughout the day increases the risk that a late-afternoon flight be delayed. Weather conditions such as rainfalls or extreme temperature play an important role too, and finally, delays depend a lot on the air carrier selected, with delays rates varying threefold amongst the top 10 carriers.