Buildings - Special Topics in Structures

Full Session with Abstracts

339370-4 - Application of statistical learning models for efficient seismic risk assessment of large property portfolios

Saturday, April 21
8:00 AM - 9:30 AM
Location: 202A

This study presents a computational framework to significantly reduce the required computational resources for portfolio risk assessment against natural catastrophes in property insurance and reinsurance applications. These applications often involve evaluating the aggregated losses to a building portfolio from a set of stochastic hazard events –such as earthquakes. Since an insured commercial portfolio in practice spans over a large geographical area, the number of seismic scenarios which need to be considered can be in the order of hundreds of thousands to millions, making seismic risk evaluations computationally challenging. Moreover, aggregated portfolio loss evaluation for each event has to consider the correlations among building responses which adds to the computational complexity of risk assessment. These challenges sanction portfolio risk assessments for large portfolios very difficult to manage with personal computers despite the widespread availability of large memory, multi-core PCs. Even when high performance computer clusters are employed, benchmark tests on a portfolio of 10,000 structures in California showed an average of 12 to 19 hours simulation time on the OpenQuake platform – a public package for seismic risk assessments. To improve the computation time, most commercial packages for property insurance use a fixed, portfolio-independent stochastic event set which contains significantly fewer events than the original set. Although the reduced set is optimized, it may miss non-negligible events for a number of portfolios, depending on their spatial distribution. This study proposes an alternative framework which does not require a reduced stochastic event set, but rather improves the computational complexity per event evaluation. Event losses are estimated by employing a non-linear regression model which is developed for the portfolio in hand by statistical learning models such as Random Forests and Support Vector Machines. Accordingly, event loss computations reduce to an equivalent closed-form solution which is orders of magnitude faster to evaluate. In exchange for the improved efficiency, the statistical model produces prediction errors in event loss evaluations which are shown to be manageable in the context of portfolio risk assessment. After presenting the methodology, this paper shows its application to a 1,000-building, spatially-distributed portfolio in California. The loss exceedance probability curve and computation time from the application of the developed statistical model are compared against the same from direct simulations in OpenQuake. The presented statistical framework is an alternative to the state-of-practice in commercial catastrophe models and may offer superior performance for certain portfolios. Moreover, the presented framework can be used in conjunction with the event set optimization methods to further enhance the computation efficiency of risk assessment applications.

Keivan Rokneddin

Senior Research Scientist

Dr. Rokneddin is a Structural engineer, Statistician, and Catastrophe model developer for property and casualty insurance. His professional areas of interest include structural reliability, seismic risk assessment, Statistical (machine) learning and technical pricing. He is also interested in highway transportation networks and the application of geospatial analysis and Bayesian Statistics to property insurance products.


Send Email for Keivan Rokneddin


339370-4 - Application of statistical learning models for efficient seismic risk assessment of large property portfolios


Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for Application of statistical learning models for efficient seismic risk assessment of large property portfolios