Predicting Delays at LAX

Max Ladegaard

Student thesis: Master thesis


This paper uses different machine learning algorithms to investigate how flight delays can be predicted at Los Angeles International Airport (LAX). LAX is a large international airport located around 30 kilometers from Downtown Los Angeles. It serviced more than 80 million passengers in 2019, which made it the second busiest airport in the United States. Like most other large international airports, LAX also struggles with flight delays and how to limit their negative impact. One of the potential solutions is to use machine learning to predict which flights are going to depart on time and which flights are going to be delayed. We built three models using three different machine learning algorithms. The goal of the models was to predict whether a flight would belong to one of four delay categories: on-time, small delay, medium delay, or large delay.
It was found that we could train a Neural Network model with an accuracy score of 95%. This was our most accurate model as it beat out a Random Forest model and an XGBoost model, which had accuracy scores of 89% and 92% respectively. In the analysis, it was discovered that weather was less important a factor than anticipated. Except for the adjusted XGBoost model, the weather features have a limited impact on the output of the models. At the other end of the spectrum, the flight's airline was a surprisingly important feature. Similarly, the time of departure was consistently one of the most essential features of the models.
In terms of practical uses, we found that a flight delay prediction model can be used to enhance operational efficiency, improve passenger experience, and optimize pricing strategies. Furthermore, there is a vast potential for using machine learning and artificial intelligence in other airport domains in the future as the industry needs efficient and scalable solutions.

EducationsMSc in Business Administration and Data Science, (Graduate Programme) Final Thesis
Publication date2023
Number of pages61
SupervisorsRobert J. Kauffman