Recalibration strategies for real-time crash prediction: adaptive drift detection vs. cumulative retraining under rare-event class imbalance

Accurate crash prediction is vital for road safety, yet sustaining model performance remains challenging due to evolving traffic conditions. Despite advances in predictive modeling, recalibration has received limited attention in crash prediction research. This study addresses this gap by evaluating six recalibration strategies: static, period-aligned, cumulative, and three adaptive drift detection approaches (ADWIN, Adaptive Random Forest, and KSWIN), applied to XGBoost, AdaBoost, Random Forest, and NNet using Automatic Vehicle Identification data from Santiago, Chile (2018–2024). Class imbalance is addressed through SMOTE-based oversampling, and hyperparameters are tuned via Bayesian optimization with Optuna. Results show XGBoost and AdaBoost offer the best tradeoff between accuracy, calibration, and computational cost. Cumulative retraining emerges as most effective, as crash prediction’s rare-event nature benefits from maximizing historical data. While adaptive drift detection is theoretically appealing, its performance suggests further refinement is required for rare-event contexts. These findings offer the first systematic comparison of recalibration strategies in real-time crash prediction.