MACHINE LEARNING CAPSTONE · 2018 US FLIGHTS

Flight Arrival
Delay Intelligence

Two separate problems on one shared split: whether a flight lands 15+ minutes late, and how many minutes early or late it arrives. In both, the simple baseline won.

5.69Mraw flight records
8models compared
96.49%test ROC-AUC · classification
7.29 minmean absolute error · regression
Classification
Will the flight land 15 minutes late or more?ArrDel15
Best of 4

Logistic Regression

The baseline beat every more complex classifier on test ROC-AUC and generalized the tightest.

Test ROC-AUC
96.49%
Train–val gap
0.011
Regression
How many minutes early or late will it arrive?ArrDelay
Best of 4

Linear Regression

The baseline won outright on R² and mean absolute error; extra flexibility bought nothing.

Test R²
95.79%
Test MAE
7.29 min
Both winners are the baselines. The product is a post-departure, wheels-off update because actual TaxiOut is part of the approved feature set.