It has been called to attention that road traffic crashes (RTCs) have a major social and economic impact, but risk of accident may be analyzed at an individual level of individual isolated occurrence. Eventually we realized that there is a significant trend behind this perception: Risk is not uniform. At certain places, it appears that the severity of accidents is (constantly) larger, even when a single crash appears to be identical. Then the observation at the latter resulted in the need to have a regional accident risk assessment score that is not pegged on a single event but a more broad environmental and demographic features. This research is aimed at the development of predictive risk models of accidents at the postal-code-level by considering concurrently the historic events of accidents, population compositions and road network characteristics. We do not position this as a pure classification task, but rather estimate the severity of the accident using ensemble machine learning models, including Random Forest, XGBoost, LightGBM, and CatBoost, and take the sum of the predictions to form a numerical and accident risk score in each postal code This aggregation-based idea has been shown to be more stable in locations with uneven accident frequency, which should be useful in addressing a region with uneven accident frequency. Regression-based metrics of model performance are used to compare model performance meaningfully. The reason why the system is not being turned into a black box is to explain the importance of features and how factors are comprehended through explainable machine learning methods.density of population and road network affect risk in regions like population density and road structure. Interestingly enough, features related to infrastructure had a tendency to dominate the simple count of accidents of purely historical nature. Altogether, this paper shows that the practical and interpretable method of accident risk analysis can help to assess the safety of the region and make the informed decisionmaking process, being based on the conditions of real-world data. Stable regional risk patterns have been shown in experiments on over 49,000 postal codes, even though there is only slight prediction accuracy on an accident level.