Proceedings of the 2020 Winter Simulation Conference K.-H. Bae, B. Feng, S. Kim, S. Lazarova-Molnar, Z. Zheng, T. Roeder, and R. Thiesing, eds. EFFICIENT RISK ESTIMATION USING EXTREME VALUE THEORY AND SIMULATION METAMODELING Joseph J. Kennedy Armin Khayyer Alexander Vinel Alice E. Smith Industrial and Systems Engineering Auburn University Shelby Center Auburn, Al 36849, USA ABSTRACT This paper considers a new approach for constructing metamodels for capturing tail behavior in stochastic systems, e.g., simulation outputs. Specifically, we are concerned with the problem of global estimation of the conditional value-at-risk (CVaR) surface, given (stochastic) responses from a collection of design points. The approach combines stochastic kriging, which has previously been shown to work well for metamodeling of discrete-event simulation output, with extreme value theory, which is a powerful statistical tool for estimating tail behavior. We present the general methodology and promising results of preliminary computational experiments. 1 INTRODUCTION Executing and analyzing high-fidelity simulation models covering a wide range of parameters are often quite computationally expensive, leading to significant attention paid in the literature to the area of metamodeling, or surrogate modeling, for simulation. Applications that are concerned with characterizing risk are particularly in need of such efforts. In these applications, the decision-makers are interested in accurately describing tail behavior for the underlying distributions, which usually requires proportionally more computational effort to observe sufficient rare events. This paper proposes a metamodeling approach for globally characterizing a particular tail measure, Conditional Value-at-Risk (CVaR), combining two relevant methodologies: stochastic kriging (SK) and the peaks-over-threshold (POT) method for tail estimation. Stochastic kriging has been shown to be a powerful modeling tool for predicting discrete-event simulation outputs, including forecasting tail measures (Chen et al. 2012). While Chen et al. (2012) demonstrated promising results, their methodology is based on empirical (nonparametric) estimators of risk measures. An alternative approach to characterizing tail behavior has been studied under the umbrella of extreme value theory (EVT). Widely used in many computational areas, it so far has received rather limited attention in the operations research community in general and simulation applications in particular. EVT is known to improve tail estimates for various stochastic systems (McNeil and Saladin 1997). Consequently, in this paper, we are interested in evaluating whether this improvement persists if used within a metamodeling framework, such as stochastic kriging. To this end, we propose a two-stage metamodeling framework, which employs the peaks-over-threshold (POT) method from EVT to estimate CVaR for a set of given design points on the response surface and then applies stochastic kriging to the estimated values to enable global predictions. 385 978-1-7281-9499-8/20/$31.00 ©2020 IEEE