Avinor Airport Concurrency Prediction
Data Scientist & ML Engineer
Comprehensive airport concurrency forecasting system. Analyzed 465k+ training records spanning 7 airport groups, engineered lagged traffic features, validated with temporal splits, and deployed a calibrated ML classifier achieving superior performance across all airport groups.
View live project ↗Challenge
Avinor needed to forecast when AFIS controllers would handle multiple simultaneous aircraft communications across seven Norwegian airport groups, enabling proactive resource planning and safety protocols.
Approach
Built a two-tier prediction pipeline: baseline rate tables (group-hour and group-weekday-hour) with hierarchical fallback logic, then trained Logistic Regression and HistGradientBoosting models using engineered features (seasonal flags, planned flights, lagged concurrence) with strict no-leakage validation splits.
Outcome
Delivered a HistGradientBoosting model achieving 0.9731 AUC and 0.0503 Brier score, outperforming baselines and providing per-airport predictions for October 2025 scheduling.
Highlights
- HistGradientBoosting: AUC=0.9731, Brier=0.0503
- Per-airport AUC ranging 0.94-0.99 (groups A-G)
- Hierarchical baseline with smart fallback logic
- Strict temporal split preventing data leakage