Predicting Loss IMPLICATIONS OF CHANGES IN RAINFALL FOR FLOOD INSURANCE.

Predicting LossIMPLICATIONS OF CHANGES IN RAINFALL FOR FLOOD INSURANCE

Floods Are Bad…

Floods are consistently the costliest disaster in the US each year

In the past 5 years all 50 states have experienced some level of flooding

Between 2011 and 2013 FEMA spent $55 billion on flood relief and recovery

Homes in special flood hazard areas (SFHA) are more likely to be damaged by a flood than fire

National Flood Insurance Program (NFIP)

Created to fill the gap left by private insurers leaving the market

Intended to reduce the burden of floods on taxpayers by creating an insurance system rather than strictly disaster relief

Questions and Goals What is the relationship between flood damage depend on rainfall amounts?

Can predicted changes in rainfall be used to predict changes in flood insurance?

Floods are 'acts of God,' but flood losses are largely acts of man.

-Gilbert White

Human Data Sources and Challenges

FEMA policy and loss statistics◦ FEMA releases data on a monthly basis◦ Loss statistic are for the whole program◦ Policy data is for the month◦ Only one month is available at a time

NCAR NWS Flood Reanalysis◦ State level data, low spatial resolution◦ Yearly statistics, low temporal resolution

Physical Data Sources Three day annual maximum precipitation from NOAA COOP stations

◦ Data courtesy of Cameron Bracken◦ Dataset goes to 2013

Flood Insurance Losses

Data Structure

Total Payments Max PrecipitationPrecip Anomaly

Visualizing Losses 2012 - WWA

Visualizing Losses 2012 - CA Max PrecipTotal Payments Precip Anomaly

Steps 1. Model likelihood of a loss

2. Model the size of loss

Methods Determining likelihood of loss

◦ Logistic regression◦ CART◦ Multinomial◦ Cluster analysisDistribution fitting for loss◦ GLM◦ GPP

Logistic Regression-2012

Model Brier Score Climo

Logistic 2012 .07 .11

Logistic 2013 .25 .298

Logistic Predictions

Loss 2012Residuals 2012Payments 2012

Loss 2013

Modeling Loss-Gamma

Modeling Loss-GPD

Future Work Expand analysis further both temporally and spatially Include mitigation measures Decrease spatial scale Development of a non-stationary GPD to incorporate changes in precipitation patterns