COLLOQUIUM OF THE COMPUTATIONAL MATERIALS SCIENCE CENTER AND THE DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES – Robust Estimation of Value-at-Risk for Quantitative Risk Management: Applications to Climatology, Insurance, Accidents and Financial Analysis – Sabyasachi Guharay
COLLOQUIUM OF THE COMPUTATIONAL MATERIALS SCIENCE CENTER
AND THE DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
(CSI 898-Sec 001)
Robust Estimation of Value-at-Risk for Quantitative Risk Management: Applications to Climatology, Insurance, Accidents and Financial Analysis
U.S. Department of the Treasury
November 7, 4:30 pm
Exploratory Hall, Room 3301
Establishing robust quantitative metrics which allow decision makers to determine the amount of risk in a system with extreme loss events is a problem of interest in many scientific fields. One of the fundamental metrics which is universally accepted in all fields of risk management is the quantity known as Value-at-Risk (VaR). A subfield of risk management, modern Operational Risk Management (ORM), closely investigates methodologies on robustly estimating VaR, “Robust Estimation of VaR.” Currently, academic researchers and industry practitioners are actively looking at ways to make this estimate more statistically robust and accurate with minimal assumption requirements.
In this talk I will present two new quantitative approaches for estimating VaR that are agnostic regarding the relationship between frequency and severity: (1) Data Partition of Frequency and Severity (DPFS) using K-means to estimate VaR; (2) Distribution based partitioning (DBP) of frequency and severity using copulas. Verification is conducted on five simulated scenario datasets while validation is conducted on five publicly available datasets from four different domains: –US Financial Indices data of Standard & Poor’s 500 and Dow Jones Industrial Average; –Chemical Loss spills as tracked by the US National Coast Guard; –Australian automobile accidents; –US hurricane data. It is observed that previous VaR calculations inaccurately estimate the VaR for 80% of the cases in simulated data and 60% of the cases in real-world data studies while new methodologies attains accurate VaR estimates which are within the 95% confidence interval bounds for both simulated and real-world data.
Refreshments will be served at 4:15 PM.
Find the schedule at http://www.cmasc.gmu.edu/seminars.htm