Traffic Incident Impact Prediction Using Machine Learning and Explainable AI: Evidence from Istanbul

Traffic incident impact prediction remains challenging for intelligent transportation systems due to complex spatiotemporal dependencies. This study analyzes 38,430 real-world traffic incidents from Istanbul (2022–2024) to predict normalized traffic deviation ΔTraffic(%) using machine learning with rigorous temporal validation. Three models—Random Forest (RF), XGBoost, and LightGBM—were evaluated using rolling-origin cross-validation (2022 training, 2023 testing; 2022–2023 training, 2024 testing) to prevent temporal leakage, employing a strictly operational 13-feature set that excludes information unavailable at incident onset (t0). LightGBM achieved MAE = 26.81 ± 1.94% and R2 = 0.506 ± 0.042 (mean ± std across folds) with 95% bootstrap confidence intervals of [27.54%, 28.81%] for MAE on the 2024 test set, significantly outperforming historical baselines (R2 = 0.100 ± 0.054, p < 0.001, Bonferroni-corrected). Feature ablation studies revealed that temporal features contribute 65.2% of predictive power, while incident type contributes only 1.3%. Distributional robustness analysis confirms conclusions are stable across distributional treatments (log, winsorised, quantile), with feature importance rank correlations ρ = 1.000 between all treatment pairs. This work provides empirical evidence for context-aware traffic management systems and demonstrates the importance of proper temporal validation in transportation forecasting.