A Simulation Analytics Approach to Dynamic Risk Measurement

Project: ResearchGRF

View graph of relations


Financial risk measures are important indicators of risk of financial portfolios and even stability of financial institutions. Therefore, it is critical to be able to estimate them accurately and promptly. Estimating risk measures is typically a challenging task due to the complicated structures of securities contained in the portfolios. Previous work in the literature focuses almost exclusively on estimating risk measures in a static way, and considers the estimation problem only once. Therefore, almost all methods are designed to be implemented over night or weekend. They provide accurate estimates of risk measures, but cannot be applied in real time. Taking advantage of the recent development of statistical learning methods and big data analytics, we propose to take an alternative view and formulate the dynamic risk estimation problem as a statistical learning problem using simulated data. This approach is generally known as simulation analytics and has been proposed only very recently. Our goal is to provide fast and accurate estimates of risk measures so that they can be used to monitor portfolio risks in real time. In this project we propose to build regression models (e.g., linear regression models or logistic regression models) using data generated in past simulation experiments and to use the models to predict portfolio risk measures and conduct risk monitoring or control. We also explore statistical learning tools, such as regularization, to improve the prediction accuracy by better balancing the tradeoff between bias and variance and to select important risk factors for dynamic risk monitoring. The statistical learning based methods can in general be viewed as black-box methods. They conduct data mining on the simulated data without exploring the information of the simulation model itself. However, in the field of risk management, analysts often possess a lot of information on the simulation model itself, and we propose to use this information to further improve the prediction quality of the regression model. To do that, we propose to consider several issues, including how to manipulate the simulated sample paths so that they appear in important regions, how to use neighborhood time information, and how to incorporate stylized models into basis functions of regression models, and evaluate these enhancement methods both theoretically and empirically.


Effective start/end date1/01/18 → …