Predicting taxi-out time of departing flights

A CatBoost pipeline trained on 12 months of EUROCONTROL ADR data — ~2.1 million departures — for the EUROCONTROL PRC Data Challenge 2026.

Final RMSE 288.75 Rank 108 / 192 CatBoost · Kaggle GPU 35 features · 7 models

RMSE improvement

−52.4%

607.15 → 288.75

Leaderboard climb

+68

rank 176 → 108

Training departures

≈2.1M

12 months · 2025

Submissions

best one counts

Score journey — lower is better

607.15

288.75

First submission

Final submission

July 2025 fold (6 months of training data)

Baseline median

719.99

Model

661.82

Dec 2025 fold (11 months of training data)

Baseline median

487.15

Model

431.42

■ Flight & airport (10)

■ Schedule & timing (10)

■ Congestion (10)

■ Historical stats (5)

EGLLLondon Heathrow LFPGParis CDG EDDFFrankfurt EHAMAmsterdam LIRFRome Fiumicino LTFMIstanbul

EGLL — one of the busiest hubs in the set; its dedicated model captures local runway-layout and stand congestion patterns.

LFPG — multi-terminal layout with distinctive stand-to-runway distances; strong beneficiary of the airport-specific model.

EDDF — high departure density per runway-hour makes congestion features especially informative here.

EHAM — heavy coordinated scheduling; queue-position features carry strong signal.

LIRF — seasonal peaks dominate; historical month-by-month statistics help most here.

LTFM — newer megahub with growing traffic; the dedicated model tracks its rapid ramp-up.

Each hub gets its own CatBoost model; predictions are blended 70% airport / 30% global.

1 · ADR data
12 monthly files, DEP rows, raw taxi-out target

2 · Features
Flight, schedule, congestion, historical stats — no post-off-block leakage

3 · Models
1 global CatBoost + 6 airport models, GPU, 900 trees

4 · Blend & submit
70/30 hub blend, validated with expanding-time folds