Forecasting Fatal Traffic Accidents During Early Covid-19 And Increased Enforcement Activities: A Case Study from Hatay Province, Türkiye Using Classical Time Series and Ensemble Machine Learning Models

This study investigates the temporal dynamics and predictive modelling of fatal traffic accidents in Hatay province, Türkiye, using both classical time series approaches (ARIMA, SARIMA, Holt–Winters) and machine learning techniques (Random Forest, Gradient Boosting). Monthly accident data from 2017–2021 were analysed through seasonal decomposition, stationarity testing, and comparative model evaluation. Results revealed a distinct seasonal pattern, with accident counts peaking during summer months and declining in winter, and a long-term trend showing a notable reduction in fatalities after 2017. Among the tested models, the Enhanced Gradient Boosting approach demonstrated the highest predictive accuracy (R² = 0.97, RMSE = 1.59), outperforming both classical time series and other ensemble methods. Forecast results for 2021 indicated seasonal peaks in June and August, corresponding to increased traffic density during the holiday period. The COVID-19 pandemic was associated with a marked short-term reduction in fatalities, though the effect appeared to diminish post-lockdown. These findings highlight the value of integrating advanced ensemble learning methods into traffic safety forecasting and underscore the importance of seasonally targeted interventions.