Accidents in transportation sector are a major public safety concern, particularly in high-traffic urban regions where increasing vehicle density and complex road conditions contribute to frequent crashes. The capacity to deliver timely early warnings based on predictive risk analysis is severely lacking in traditional traffic management systems, which are mostly reactive. This paper presents a solution to this constraint by leveraging the US Accidents (2016-2023) dataset, which includes 2,845,342 accident records with 46 characteristics, and proposing a prediction model based on Extreme Gradient Boosting (XGBoost) to provide early warning of traffic accidents. The proposed method incorporates exploratory data analysis, data preparation, label encoding, min-max normalization, SMOTE-based class balancing, and an 80:20 training-testing split, among a number of other things. The XGBoost model uses accuracy, precision, recall, and F1-score to evaluate the efficacy of its accident risk forecasting training. With an accuracy (ACC) of 95.9%, precision (PRE) of 96.0%, recall (REC) of 95.7%, and F1-score (F1) of 95.8%, the proposed model surpasses state-of-the-art machine learning (ML) methods, test results show, including FL, Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM). According to these results, the XGBoost model is a solid and efficient way to anticipate when traffic accidents happen, which helps with IVSs by allowing for faster risk assessments and safer roads.