Application of Statistical Learning Models for Efficient Seismic Risk Assessment of Large Property Portfolios
Rokneddin and Shahjouei
Seismic risk assessments are essential for the property and casualty insurance and many public entities at the national and local levels; however, comprehensive studies for large portfolios and in seismically active regions often become time-consuming processes and require significant computational resources to run the required simulations. This research introduces surrogate models which are developed by random forests, a class of nonlinear statistical learning algorithms, to significantly reduce the computational requirements in exchange for manageable errors in predicting the portfolio losses. To demonstrate the application, a portfolio consisting of four different building classes is simulated in OpenQuake, an open-source platform for seismic risk analysis. The developed surrogate model is shown to save close to 70% of the computation time and predict the portfolio losses for small to mid-range events with small errors but underestimates the values for very large events. A similar method can be applied to develop parametric solutions.