CyberRisk Prediction using Machine Learning and Extreme Value Theory

This paper develops a hybrid framework for quantifying the financial impact of data breaches by combining predictive machine learning with extreme value theory (EVT). Using incident-level breach data from the Privacy Rights Clearinghouse (PRC) covering the period 2005–2020, we first estimate the number of compromised records with a Random Forest model trained on organizational, temporal, and attack-type characteristics. We then analyze the tail behavior of the predicted losses to capture the fat-tailed distribution of cyber risks. Our results indicate that the distribution of affected records is well represented by a Fr´echet law, and we estimate the parameters of the Generalized Extreme Value (GEV) distribution to compute Value-at-Risk (VaR) at high confidence levels. This two-stage approach provides a rigorous assessment of maximum potential losses, addressing the question of cyber-risk insurability. By linking predictive accuracy with tail risk quantification, our findings deliver actionable insights for insurers, regulators, and organizations seeking to anticipate and manage the f inancial consequences of large-scale data breaches.