Combining a novel metric and categorical boosting to better estimate rare events likelihood: Use in traffic accidents

Recent data indicates a notable increase in urban accidents, particularly in Spain. Numerous factors contribute to accidents, often involving various combinations of them. This study aims to assess the probability of an accident occurring under specific conditions in Madrid. Understanding the probability of accidents under specific conditions is crucial for improving road safety, optimizing traffic management, and informing policy decisions. We apply Categorical Boosting (CatBoost) to estimate accident probabilities using three complementary approaches. The first approach utilizes all predictor variables and the entire dataset. The second approach removes the road variable and estimates probabilities separately for each road, assuming that road-specific models might improve prediction quality. The third approach excludes time-related predictor variables, creating models, one per hour, to assess whether time is the most influential factor. To evaluate model performance, we introduce a novel metric quantifying how much the probability of an accident increases when one has already occurred. Additionally, we use traditional metrics. Our findings demonstrate that CatBoost models consistently outperform baseline models in most cases, achieving higher accuracy and balanced accuracy scores on average. Therefore, the results highlight the effectiveness of CatBoost in predicting accident probabilities and suggest that road-specific and time-specific modelling approaches can provide valuable insights for urban accident analysis and prevention strategies. • New CatBoost models for predicting the probability of accidents. • Our models significantly improved handling of categorical variables. • A new metric to properly deal with imbalanced scenarios. • Our CatBoost models, based on our metric, consistently outperform baseline models.