Disaster managers rely heavily on historical data obtained from previous disasters to predict the severity of future disasters and their potential risks to people living in the area. This paper proposes a general framework to predict the severity of cyclones based on a number of injuries and fatalities reported, as well as identifying high-risk areas and estimating the fatalities during cyclones. To demonstrate the framework’s effectiveness, we use historical data obtained from cyclones in Mozambique between 2017 and 2025. Our framework integrates classification and regression techniques to accurately model cyclone impact. The result is an F1-score of 0.872 for random forest, which estimates cyclone severity prediction, and 0.812 as low absolute error of support vector regression to estimate fatalities. High-risk zones were identified by aggregating severity probabilities at district and provincial levels, closely aligning with historically affected regions like Sofala and Zambezia. These results demonstrate the feasibility of using ML for more precise impact prediction, enhancing the ability of disaster managers to allocate aid and plan interventions more effectively.