<p>This study applies state-of-the-art machine learning (ML) classifiers and optimization techniques in the development of a unique framework to forecast the probability and effect of cybersecurity attacks. This framework includes the GPC and Stacking Classifier, which improve the forecast accuracy by stacking a number of base models due to their capacity to represent intricate patterns and uncertainties in cybersecurity data, meaning the model can learn complex, nonlinear relationships between attack features and outcomes, while also estimating prediction confidence through probabilistic outputs. This is particularly useful for cybersecurity, where attack behaviors are often dynamic, ambiguous, and context-dependent. These classifiers are optimized using two cutting-edge algorithms: Northern Goshawk Optimization (NGO) and Bald Eagle Search Optimization (BESO), which guarantee better hyperparameter tuning and convergence. Based on a comprehensive comparison using accuracy, precision, recall, and F1-score across training and testing phases, the STBE model achieved the highest performance with an overall accuracy of 0.992, while the GPC model demonstrated the weakest performance with an overall accuracy of 0.965, marking STBE as the most effective and GPC as the least effective model in the experimental framework. A filter-based feature selection algorithm, the f-statistic (f_classif) method, has been utilized to reduce computation complexity while preserving essential predictive information. The process assesses the statistical difference of each feature from the target variable to enable the extraction of essential indicators of cyber attacks, like system vulnerabilities and network anomalies on the basis of their discrimination power.</p>