US Hurricanes and economic damage: an extreme value perspective Nick Cavanaugh, futurologist Dan Chavas, tempestologist Christina Karamperidou, statsinator Katy Serafin, bathy queen Emmi Yonekura, landfaller ASP 2011 Summer Colloquium Project 23 June 2011
Outline Motivation Previous work Methodology and results – Economic data: absolute vs. relative damages – GPD without physical covariates – GPD with physical covariates – Application to GFDL current vs. future hurricanes Conclusions and future work
Motivation: society Atlantic hurricane tracks (1900+) (NHC Best Track) GDP: 1 o x 1 o (Yale G-Econ) 63% of global insured natural disaster losses caused by US landfalling hurricanes (Source: Rick Murnane, last week)
Motivation: science Objectives: – Combine physical storm characteristics with statistics of damages in an extreme value theory framework – Reduce the sensitivity of statistical analysis of damage to economic vulnerability at landfall
Recent work Katz (2002), Jagger et al (2008,2011) Jagger et al (2008,2011): Generalized Pareto Distribution (GPD) is appropriate for modeling extreme events involving large economic losses However, inclusion of physical characteristics of storms as covariates has not been tried
Methodology I: absolute vs. relative damage Economic data: Pielke et al., 2008 Base year and normalized (2005$) economic damages for 198 storms (pre-threshold) from But are variations in damages representative of the damage threat from a hurricane or rather of the large variation in economic value along the coast? Distribution of GDP (bil $) in 1 o x 1 o boxes along US coast
Methodology I: absolute vs. relative damage Damage Index (DI) Fraction of possible damage [0,1] i.e. “damage capacity” of storm Economic Physical Goal: remove from our damage database the variability in damages due to variations in economic value along the coast Physical characteristics of storms and economic value at landfall should be independent corr = -.1 Neumayer et al. (2011) *
Histogram of Total Damage:Histogram of Damage Index: Results Damages vs. DI: histograms Max = $150 bil Max =.89
Total Damage: (bil 2005$) Damage Index (DI): [0,1] Great Miami $156 bil Bret.89 Top 10 by Damage: Top 10 by DI: Results Damages vs. DI: no covariates
Results Damages vs. DI: no covariates Total Damage: (bil $)Damage Index (DI): [0,1] ξ > 0ξ ~ 0
Results Damages vs. DI: no covariates Total Damage Damage Index (DI)
Methodology II: physical covariates Want to capture physical characteristics of individual storms that are relevant to its capacity to cause damage
Hurricane Katrina 8:15p CDT Aug
Hurricane Katrina 8:15p CDT Aug Eye
Hurricane Katrina 8:15p CDT Aug Eyewall
Hurricane Katrina 8:15p CDT Aug R 34
Methodology II: physical covariates Wind Storm surge Sensitive to: -Wind speed (V max ) -Size (R 34 ) Sensitive to: -Wind speed (V max ) -Size (R 34 ) -Bathymetry (s eff ) -Translation speed -Landfall angle Causes of damage See Irish et al. (2008)
Methodology II: physical covariates Wind speed V max : HURDAT Best Track Storm size R 34 : Extended Best Track (CSU) Bathymetry: gridded 1-min res altimetry data 100 km s eff
Methodology II: physical covariates Bathymetry
Methodology III: GPD fit PDF With Multiple Possible Covariates
Results Damage: with covariates Damages *Using data r 34 : not enough data shape parameter left constant Damage = f(V max )
Damage Index *Using data Results DI: with covariates DI = f(s eff, V max ) r 34 : not enough data shape parameter left constant Likelihood -ratio test
Methodology IV: Future Climate Statistical-Deterministic Hurricane model (Emanuel et al. 2006) – downscaled from GFDL CM2.0 model: and (A1b) climates Modeled values of V max and s eff => GPD
Results: Future Climate GPD PDF of US Hurricane Damage Index Add all PDFs and re-fit GPD for each climate
Results: Future Climate Local Distribution of Scale Parameter Change Δσ local =Δ exp( σ 0 + σ 1 V max + σ 2 s eff )
Conclusions Damage Index, which seeks to remove economic vulnerability from damages, appears to better capture role of physical characteristics of storm in causing damage than actual damages Bathymetry, wind speed found to be useful covariates whose relationships are consistent with physical intuition Changes in scale parameter in the future indicate a shift to higher probability of extreme damage events locally and globally, though we haven’t proven differences are statistically significant
Future work ideas Find means of relating back to actual economic damages Try r max for size Account for uncertainty Try out a deterministic damage index and apply GPD to that? Thanks! Comments/suggestions welcome
Results Damages vs. damage index DI = f(s eff )
Results Damages vs. damage index DI = f(V max )
Results: Future Climate
Top 10 by Wind Speed:
Example 1: Katrina vs. Camille Peak storm surge = 8.5 m Peak storm surge = 6.9 m NOAA SLOSH model KATRINA (2005)CAMILLE (1969) …yet Katrina produced much higher storm surge because it was twice as large