Analysis and Uncertainty Reduction through Data Augmentation in Hurricane Risk Modeling in the State of Florida
Otarola Farah, Daphne Raquel
MetadataShow full item record
The Florida Public Hurricane Loss Model (FPHLM) and the Florida Public Flood Loss Model (FPFLM) project insurance losses due to hurricane events in the State of Florida from hurricane induced wind, rain, and flood. This hurricane catastrophe model estimates annual average loss costs, as well as mean loss levels for scenario events, for personal and commercial residential portfolios, including building structure, content, appurtenant, and additional living expenses. Insurance companies use these hurricane risk models in Florida, and in particular the Florida Office of Insurance Regulation (OIR) mandates that the FPHLM be used for the stress test, that verifies the actuarial soundness of the companies. Catastrophe models like the FPHLM/FPFLM estimate risk at the intersection of hazard, exposure, and vulnerability. Each of these areas requires diverse sources of data, which are very often incomplete, inconsistent, or missing altogether. For example, to define the input exposure, and for model development, calibration, and validation of the FPHLM and the FPFLM vulnerability curves, the FPHLM teams accessed three different data sources: county tax appraiser (TA) databases, National Flood Insurance Program (NFIP) portfolios, and wind insurance portfolios. The poor quality of these data is a source of epistemic uncertainty, which affects the vulnerability models as well as the output of the catastrophe models. This thesis presents an extensive exposure and claim data study for the purpose of reducing this uncertainty and improving the quality of the data through identification, augmentation, and integration of the different databases. The reformatting, processing and matching of the insurance databases with the TA data resulted in more accurate information in: 1) the exposure portfolio, which are the input to the FPHLM/FPFLM; 2) the claims portfolios which are the basis for development, validation and calibration of the vulnerability curves. Finally, this augmented data resulted in a new set of improved statistics on building characteristics for the overall Personal Residential (PR) and Commercial Residential (CR) properties in Florida. These statistics provide an accurate description of the building population in Florida, and they are the basis for the weighting of the vulnerability functions for loss analysis, and the random assignment of missing parameters in the insurance portfolios. In conclusion, the augmented wind and NFIP data impact the FPHLM and FPFLM in three different ways: 1) it improves the quality of the input data, and therefore the output of the FPHLM and FPFLM; 2) it improves the development and validation process of the vulnerability model, therefore improving further the quality of the loss analysis output; 3) it results in more accurate building population statistics, which results in a better weighting of the vulnerability function and a more accurate random assignment of missing parameters in the insurance portfolios. Overall, the impacts on the uncertainty reduction could be significant and contribute to better results. The lessons learned can be extended to other models and other hazards.