On a trip back to China, Naitian was stuck at SFO for a solid 14 hours. It was kind of terrible.
What it does
Given a flight number, it'll predict how early or late your flight will arrive.
How we built it
We found a dataset of flight arrivals spanning from 1987 to 2018. We used Azure's machine learning studio to both explore the dataset as well as to train a linear regression model to predict flight delays. Then we used Wix Code to build a useable front end that calls our API.
Challenges we ran into
Wrangling the large dataset. Each month has almost half a million entries.
Accomplishments that we're proud of
We learned how to use Azure for machine learning, experimentation, and inference in production.
What we learned
Flights are actually more on time than we expected. The average delay was around 7 minutes, but if we look at the median delay (which accounts for the non-gaussian distribution of delay times), flights are actually generally early (by 3 minutes).
What's next for Late to the Gates
Incorporating weather data and other external features to improve the accuracy of our model.