The impact of COVID-19 on machine learning models in a commercial aviation use case

The COVID-19 pandemic has led to unforeseen changes in the world. These changes have had a serious impact on many machine learning models. These models rely on historic data to generate forecasts and make decisions. The changes in society have led to these forecasts and decisions not matching reality, with potentially catastrophic results. The aviation industry is the industry that has been affected the most by the pandemic. Airliners rely on machine learning models to e.g. schedule flights and price tickets. The aim of the thesis was to see if it was possible to develop models capable of predicting airline passenger amounts, research if it was possible to measure the degradation in performance between years and discuss potential ways to fix the models. The forecasting procedures used in developing the models were Prophet and SARIMAX. The data used was the monthly amount of air passengers carried by flights in the USA between the years 2003-2018. The models were then used to forecast the years 2019 & 2020 and then compared to actual numbers. The models were capable of forecasting the year 2019 with a margin of error of only a few percent. When forecasting 2020 the forecast was off by over 60%, making it clear that the models were incapable of adapting to the changed circumstances. There proved to be various ways one could try to fix the model. Active tracking of the models followed by the introduction of new data could lead to the models making connections between old and new data, potentially making the models able to adapt to the new circumstances