Abstract
Winter departure delays impose sizable costs to airlines, about $75 per delay minute. We modeled determinants of push-back delay for 130,166 departures from Los Angeles (LAX) and Chicago O’Hare (ORD) between November 2024 and April 2025 by linking BTS On-Time records to NOAA daily precipitation and temperature, and implementing a workflow in R that combined EDA, multiple regression with interactions, and elastic-net shrinkage. The specification set comprised a pooled OLS with Airport × TaxiOut, precipitation, and temperature interactions, airport-specific OLS, and elastic-net with alpha 0.5; diagnostics indicated acceptable multicollinearity (GVIFadj < 2.5), mild heteroskedasticity (Breusch-Pagan p=0.07), heavy tails (Anderson-Darling p<0.001), and no autocorrelation (Durbin-Watson p=0.98). Mean departure delay was 13.2 minutes at ORD and 7.4 minutes at LAX with a pronounced right tail at ORD. Each additional taxi-out minute was associated with 0.19 minutes of departure delay at both airports. Rainfall sensitivities diverged markedly, with one inch adding 14 minutes at LAX versus 56 minutes at ORD. ORD exhibited Monday and Friday penalties of roughly 6 minutes relative to Tuesday and Wednesday, while LAX showed little weekday structure. The elastic-net retained TaxiOut, precipitation, and an adverse-weather by ORD effect while shrinking 51 indicators. Out-of-sample RMSE was about 21 minutes. A minimal padding policy, 0.7 minutes at LAX and 1.3 minutes at ORD, is projected to reduce total delay minutes by 10 percent. Explanatory power was low (R² 2 to 4 percent), reflecting unobserved operational constraints; results nevertheless highlight surface management and weather hardening as high-leverage levers.