This study proposes a threshold adjustment method at the output stage, without using sampling, to simultaneously address the issues of class imbalance and class overlap in traffic accident prediction. To overcome the limitations of data-level techniques such as SMOTE, which can cause data distortion or boundary expansion, the study enhances rare accident detection performance through Fβ-score-based threshold optimization. Experiments using real high-resolution spatiotemporal data (Daejeon intersections, in 3-hour intervals) demonstrated that the proposed threshold adjustment strategy significantly improved recall while relatively suppressing false alarms, achieving a balanced performance. These results suggest that the method is a lightweight and practical approach that can be effectively applied to proactive accident prevention and alert policy formulation in traffic safety operations.