APPENDIX 1 R CODE ILLUSTRATIONS

# Dag Lohmann, KatRisk LLC, June 2012 # katrisk.R explains the basics of risk modeling # The code assumes some abitrary hazard and vulnerability functions that # would normally be peril dependent (e.g. as functions of peak wind gust # or flood depth) # The code contains all features of a risk model including secondary # uncertainty # Here are the steps to run the code # 1) install R # 2) install.packages("copula","evd","fields","maps") # type source("katrisk.R") in the R window - or use RStudio to go through # the code # This is what the code does: # 1) create hazard events # 2) create exposure # 3) create vulnerability # 4) build an event loss table # 5) define and run financial model # 6) compute OEP and AEP curves for expected losses # 7) Include secondary uncertainty as beta distributed around the mean loss # 8) build an event loss table for sampled losses # 9) run financial model on sampled losses # 10) compute OEP and AEP curves for sampled losses # 11) compute other analytical quantities # # # # - loss cost - probable maximum loss at 250 year RP - contribution of AEP return period losses to AAL - EP uncertainty

# load libraries, set seed for repeatability, set plot parameter library("copula") library("evd") library('fields') library('maps') set.seed(1) # 1) create hazard # This is normally done with detailed vendor specific peril models. Model 10

# # # #

output can be stored by either geographical location (e.g. lat/lon for 1 degree grid cells), or by administrative boundaries such as postcodes. The important role of geocoding is then to align the exposure to each of these.

# number of grids in the hazard field (this would be equivalent to producing # e.g. quadkeys for n grids and giving them keys # n.grids is then just the number of different possible geolocations # n.events is the number of events, each with a certain rate. These could be # non-uniform of course. rho.par is the correlation between locations. This would normally be calculated # by a hazard model produce hazard on grid that looks as if model was built for the UK # but don't care whether losses are over land or ocean x.coord <- c(-6:2) y.coord <- c(50:59) dim.hazard <- c(length(x.coord),length(y.coord)) n.grids <- dim.hazard[1]*dim.hazard[2] n.events <- 10000 rate.per.event <- 1/10000 event.rates <- rep(rate.per.event,n.events) rho.par <- 0.5 # create n.events events of some random hazard with an elliptic copula and # GPD marginal distribtions myCop.norm <- ellipCopula(family="normal",dim=n.grids,dispstr="ex",param=rho.par) gpd.par <- list(loc=1,scale=8,shape=0.5) myMvd <- mvdc(copula=myCop.norm,margins=rep("gpd",n.grids),paramMargins=rep(list(gpd.par),n.grids)) hazard <- rmvdc(myMvd,n.events) # Plot first 4 hazard fields png(filename=paste('first4hazard.png'), width = 800, height = 800) par(mfrow=c(2,2),mgp=c(2.3,0.7,0),oma=c(0,0,2,0),las=1) for (i in c(1:4)) { hazard.plot <- hazard[i,] dim(hazard.plot) <- dim.hazard hazard.plot <- ifelse(hazard.plot>100,100,hazard.plot) image.plot(x.coord,y.coord,hazard.plot,zlim=c(0,100),xlab="Longitude",ylab="Latitude",cex=2.0) map(add=T,col='white',lwd=2) } title(main="First 4 Hazard Fields",cex.main=1.7,outer=T) dev.off() 11

Illustration : First four hazard fields. In this case the hazard is random, and therefore would be replaced by a real catastrophe model. # 2) create exposure. I assume here that n.loc locations are randomly # # distributed in the n.grids hazard data set. Normally geocoding would take care of this and assign each location to a hazard grid

# number of locations in the exposure set n.loc <- 20 # create exposure, e.g. $100k per subject at risk exposure <- rep(100000,n.loc) # give each of the locations a random location in the n.grids event grids location.in.hazard <- as.integer(runif(n.loc)*n.grids)+1 # 3) Create vulnerability # a parameterized function for a hazard between 0 and 100 damage.ratio <- function(hazard) { hazard <- ifelse(hazard<20,0,hazard) hazard <- ifelse(hazard>99.99,99.99,hazard) damage.ratio <- (hazard/100)**6 } # 4) Create the event loss table 12

# this table contains by event the ground up losses for each location # create top 10 events gu.dr.loss.table <- damage.ratio(hazard[,location.in.hazard]) gu.event.loss.table <- exposure*gu.dr.loss.table gu.loss.by.event <- apply(gu.event.loss.table,1,sum) top.events <- order(gu.loss.by.event,decreasing=T)[1:4] # plot top 4 events png(filename=paste('top4events.png'), width = 800, height = 800) par(mfrow=c(2,2),mgp=c(2.3,0.7,0),oma=c(0,0,2,0),las=1) for (i in c(1:4)) { hazard.plot <- hazard[top.events[i],] dim(hazard.plot) <- dim.hazard hazard.plot <- ifelse(hazard.plot>100,100,hazard.plot) image.plot(x.coord,y.coord,hazard.plot,zlim=c(0,100),xlab="Longitude",ylab="Latitude") map(add=T,col='white',lwd=2) } title(main="Hazard Top 4 Events",cex.main=1.7,outer=T) dev.off()

Illustration : Top 4 events based on the total loss # 5) Define financial model # # # for simplicity same financial structure with deductible and limit for all exposures Compute the gross loss event loss table 13

l.deductible <- 1000 l.limit <- 80000 gr.event.loss.table<-ifelse(gu.event.loss.table<l.deductible,0,gu.event.loss.table-l.deductible) gr.event.loss.table<-ifelse(gr.event.loss.table>l.limit,l.limit,gr.event.loss.table) gr.loss.by.event <- apply(gr.event.loss.table,1,sum) # compute average annual loss (AAL) for each location gu.aal <- apply(gu.event.loss.table*event.rates,2,sum) gr.aal <- apply(gr.event.loss.table*event.rates,2,sum) # plot AAL for GU and GR loss gu.aal.plot <- rep(0,n.grids) gr.aal.plot <- rep(0,n.grids) for (i in c(1:n.loc)) { gu.aal.plot[location.in.hazard[i]] <- gu.aal.plot[location.in.hazard[i]] + gu.aal[i] gr.aal.plot[location.in.hazard[i]] <- gr.aal.plot[location.in.hazard[i]] + gr.aal[i] } dim(gu.aal.plot) <- dim.hazard dim(gr.aal.plot) <- dim.hazard png(filename=paste('aal.png'), width = 800, height = 400) par(mfrow=c(1,2),mgp=c(2.3,0.7,0),oma=c(0,0,2,1),las=1) image.plot(x.coord,y.coord,gu.aal.plot,xlab="Longitude",ylab="Latitude",zlim=c(0,max(gu.aal.plot))) map(add=T,col='white',lwd=2) title(main="GU AAL",cex.main=1) image.plot(x.coord,y.coord,gr.aal.plot,xlab="Longitude",ylab="Latitude",zlim=c(0,max(gu.aal.plot))) map(add=T,col='white',lwd=2) title(main="GR AAL",cex.main=1) dev.off()

Illustration: This figure shows the average annual loss for GU (left) and GR (right) losses # 6) Compute OEP and AEP curve through simulation of n.years 14

# # # # # #

this is also possible for some analytical distributions of frequency distributions OEP = occurrence exceedance probability, the distribution of the maximum loss per year AEP = aggregate exceedance probability, the distribution of annual (normally, but could use other aggregation periods) losses

# assume a Poisson distribution for the number of occurrences # model this with arrival times that years*U[0,1] distributed # run each event at a minimum n.sim times # only simulate with events that produce a loss # round the number of expected loss occurrences n.sim <- 10 loss.events <- which(gu.loss.by.event>0) n.years <- n.sim/min(event.rates) gu.oep.losses <- rep(0,n.years) gu.aep.losses <- rep(0,n.years) gr.oep.losses <- rep(0,n.years) gr.aep.losses <- rep(0,n.years) for (i in loss.events) { n.occur <- round(n.years*event.rates[i]) year.occur <- as.integer(runif(n.occur)*n.years)+1 for (j in year.occur) { gu.oep.losses[j] <- max(gu.oep.losses[j],gu.loss.by.event[i]) gr.oep.losses[j] <- max(gr.oep.losses[j],gr.loss.by.event[i]) gu.aep.losses[j] <- gu.aep.losses[j]+gu.loss.by.event[i] gr.aep.losses[j] <- gr.aep.losses[j]+gr.loss.by.event[i] } } # order the losses and rates ord.gu.oep.losses<- gu.oep.losses[order(gu.oep.losses,decreasing=T)] ord.gr.oep.losses<- gr.oep.losses[order(gr.oep.losses,decreasing=T)] ord.gu.aep.losses<- gu.aep.losses[order(gu.aep.losses,decreasing=T)] ord.gr.aep.losses<- gr.aep.losses[order(gr.aep.losses,decreasing=T)] rates <- c(1:n.years) / n.years # plot the results # first part of the plot is produced here, second part includes the results with secondary uncertainty png(filename=paste('ep.png'), width = 800, height = 400) 15

par(mfrow=c(1,2),mgp=c(2.3,0.7,0),oma=c(0,0,2,1),las=1) max.plot <- max(ord.gu.aep.losses) y.lim <- log10(c(1/n.years,1)) x.lim <- c(0,max.plot) plot(ord.gu.oep.losses,log10(rates),ylim=y.lim,xlim=x.lim,xlab="Loss",ylab="Return Period",las=1,type="l",lwd=2,axes=F) lines(ord.gu.aep.losses,log10(rates),col="red",lwd=2) lines(ord.gr.oep.losses,log10(rates),col="blue",lwd=2) lines(ord.gr.aep.losses,log10(rates),col="green",lwd=2) p.intervals = log10(n.years)+1 ep.plot <- 10**c(p.intervals:0) axis(1,at=pretty(c(0,max.plot),p.intervals),labels=pretty(c(0,max.plot),p.intervals)) axis(2,at=log10(1/ep.plot),labels=as.character(ep.plot),las=1) abline(h=log10(1/ep.plot),col="lightgray",lwd=2,lty=2) abline(v=pretty(c(0,max.plot),p.intervals),lwd=2,col="lightgray",lty=2) title(main="EP Results",cex.main=1.7,outer=T) # 7) Include secondary uncertainty as beta distribution around the mean loss for each loaction with a correlation # that is not dependent on location or distance. That can be changed in the copula. # define functions for beta distribution, a and b parameter, and mdr-cv relationship a.func <- function(mdr,cv)(1-mdr)/cv**2-mdr b.func <- function(mdr,a) a*(1-mdr)/mdr mdrcv <- function(mdr,p=1)(sqrt((1-mdr)/(p+mdr))) # define n.sample as the number of samples taken # easy to just use n.sim n.sample <- n.sim # define the correlation of uncertainty rho.uncertainty <- 0.1 # calculate the number of sampled events and provision results matrix and new event rates n.sample.events <- length(loss.events)*n.sample sampled.damage.ratio <- matrix(0,n.sample.events,n.loc) s.event.rates <- rep(event.rates[loss.events]/n.sample,each=n.sample) counter <- 0 for (i in loss.events) { loss.locs <- which(gu.dr.loss.table[i,] > 0) n.l <- length(loss.locs) beta.cv <- mdrcv(gu.dr.loss.table[i,loss.locs]) beta.a <- a.func(gu.dr.loss.table[i,loss.locs],beta.cv) beta.b <- b.func(gu.dr.loss.table[i,loss.locs],beta.a) if (n.l>1) { myUncertCop.norm <- ellipCopula(family="normal",dim=n.l,dispstr="ex",param=rho.uncertainty) 16

beta.par <- list() for (k in c(1:n.l)) beta.par[[k]] <- list(shape1=beta.a[k],shape2=beta.b[k]) myUncertMvd <mvdc(copula=myUncertCop.norm,margins=rep("beta",n.l),paramMargins=beta.par) mdr.sample <- rmvdc(myUncertMvd,n.sample) } else { mdr.sample <- rbeta(n=n.sample,shape1=beta.a[1],shape2=beta.b[1]) } sampled.damage.ratio[(n.sample*counter)+c(1:n.sample),loss.locs] <- mdr.sample counter <- counter + 1 } # 8) Create the event loss table for sampled events # this table contains by event the ground up losses for each location s.gu.event.loss.table <- exposure*sampled.damage.ratio s.gu.loss.by.event <- apply(s.gu.event.loss.table,1,sum) # 9) Use same financial model s.gr.event.loss.table<-ifelse(s.gu.event.loss.table<l.deductible,0,s.gu.event.loss.table-l.deductible) s.gr.event.loss.table<-ifelse(s.gr.event.loss.table>l.limit,l.limit,s.gr.event.loss.table) s.gr.loss.by.event <- apply(s.gr.event.loss.table,1,sum) # compute average annual loss (AAL) for each location # not plotted this time s.gu.aal <- apply(s.gu.event.loss.table*s.event.rates,2,sum) s.gr.aal <- apply(s.gr.event.loss.table*s.event.rates,2,sum)

# 10) Compute OEP and AEP curve through simulation as before s.gu.oep.losses <- rep(0,n.years) s.gu.aep.losses <- rep(0,n.years) s.gr.oep.losses <- rep(0,n.years) s.gr.aep.losses <- rep(0,n.years) for (i in c(1:n.sample.events)) { n.occur <- round(n.years*s.event.rates[i]) year.occur <- as.integer(runif(n.occur)*n.years)+1 for (j in year.occur) { s.gu.oep.losses[j] <- max(s.gu.oep.losses[j],s.gu.loss.by.event[i]) s.gr.oep.losses[j] <- max(s.gr.oep.losses[j],s.gr.loss.by.event[i]) s.gu.aep.losses[j] <- s.gu.aep.losses[j]+s.gu.loss.by.event[i] 17

s.gr.aep.losses[j] <- s.gr.aep.losses[j]+s.gr.loss.by.event[i] } } # order the losses and rates s.ord.gu.oep.losses<- s.gu.oep.losses[order(s.gu.oep.losses,decreasing=T)] s.ord.gr.oep.losses<- s.gr.oep.losses[order(s.gr.oep.losses,decreasing=T)] s.ord.gu.aep.losses<- s.gu.aep.losses[order(s.gu.aep.losses,decreasing=T)] s.ord.gr.aep.losses<- s.gr.aep.losses[order(s.gr.aep.losses,decreasing=T)] # Plot the results max.plot <- max(s.ord.gu.aep.losses,na.rm=T) y.lim <- log10(c(1/n.years,1)) x.lim <- c(0,max.plot) plot(s.ord.gu.oep.losses,log10(rates),ylim=y.lim,xlim=x.lim,xlab="Loss",ylab="Return Period",las=1,type="l",lwd=2,axes=F) lines(s.ord.gu.aep.losses,log10(rates),col="red",lwd=2) lines(s.ord.gr.oep.losses,log10(rates),col="blue",lwd=2) lines(s.ord.gr.aep.losses,log10(rates),col="green",lwd=2) p.intervals = log10(n.years)+1 ep.plot <- 10**c(p.intervals:0) axis(1,at=pretty(c(0,max.plot),p.intervals),labels=pretty(c(0,max.plot),p.intervals)) axis(2,at=log10(1/ep.plot),labels=as.character(ep.plot),las=1) abline(h=log10(1/ep.plot),col="lightgray",lwd=2,lty=2) abline(v=pretty(c(0,max.plot),p.intervals),col="lightgray",lwd=2,lty=2) dev.off()

Illustration : This figure shows the EP from the simulation without (left) and with (right) secondary uncertainty. 18

# 11) Define and calculate other analytical measures # # # # # # all key figures and analytical quantities can also be derived from the # simulated results with secondary uncertainty. # Here we focus on the original results (expected loss) # These computations can also be done for any loss perspective for any # aggregation level # Loss Cost is defined here as the normalized loss per 1000 exposure # It can be defined for other quantities / currencies as well loss.cost.gu.aal <- 1000* gu.aal / exposure loss.cost.gr.aal <- 1000* gr.aal / exposure cat(paste("loss cost GU / GR", loss.cost.gu.aal, loss.cost.gr.aal),fill=T) # Probable maximum loss at 250 year RP # first find the event that is the 250 year RP event # define this on GR loss loss.250 <- ord.gr.oep.losses[min(which(rates >= 1/250))] # find events that are within 5% of that loss pml.events <- which(abs(gr.loss.by.event - loss.250) < loss.250*0.05) # plot 2 of those events png(filename=paste('pml.png'), width = 800, height = 400) par(mfrow=c(1,2),mgp=c(2.3,0.7,0),oma=c(0,0,2,1),las=1) for (i in c(1:2)) { hazard.plot <- hazard[pml.events[i],] dim(hazard.plot) <- dim.hazard hazard.plot <- ifelse(hazard.plot>100,100,hazard.plot) image.plot(x.coord,y.coord,hazard.plot,zlim=c(0,100),xlab="Longitude",ylab="Latitude") map(add=T,col='white',lwd=2) title(main=paste("Loss = ",round(gr.loss.by.event[pml.events[i]])),cex.main=1) } title(main="2 PML Events",cex.main=1.7,outer=T) dev.off() - loss cost - probable maximum loss at 250 year RP - contribution of return period losses to AAL - EP uncertainty

19

Illustration : Hazard of two PML events around the 250 year return period # Contribution to AAL by return period for AEP # this can also be computed for the OEP from the events themselves # It's easy to do this straight from simulation data # # We split up the return periods into some intervals (rp.intervals) # rp.index contains from the ordered AEP vector the loss for the # corresponding return period # # the plot shows the cummulative sum of the contribution to the AAL # by return period # rp.intervals <- c(1,5,10,20,50,100,200,500,1000,n.years) rp.contribution <- rep(0,length(rp.intervals)-1) rp.index <- n.years / rp.intervals for (i in c(1:(length(rp.intervals)-1))) { rp.contribution[i] <- sum(ord.gr.aep.losses[c(rp.index[i]:rp.index[i+1])]) } rp.contribution <- rp.contribution / (sum(gr.aal)*n.years) png(filename=paste('analytics.png'), width = 800, height = 400) par(mfrow=c(1,2),mgp=c(2.3,0.7,0),oma=c(0,0,2,1),las=1) barplot(cumsum(rp.contribution),names.arg=rp.intervals[2:length(rp.intervals)], cex.names=0.8) title(main="CDF of AAL by return period",cex.main=1.7) # EP uncertainty # let's assume we have 20 years of observed loss data and want to place # these on the EP curve (can be added to plot) 20

# at the same time we can plot the uncertainty from the modeled losses # by splitting up the simulations into 20 year chunks gr.oep.losses.20 <- gr.oep.losses # create from modeled losses n.years/20 time series of losses dim(gr.oep.losses.20) <- c(20,n.years/20) # aal.20 contains the n.years/20 different modeled AALs aal.20 <- apply(gr.oep.losses.20,2,mean) # sort all of the 20 year simulations gr.oep.losses.order.20 <- apply(gr.oep.losses.20,2,sort) # Box Plots show the following (after Climate Blog): # The rectangle shows the interquartile range (IQR); it goes from the first # quartile (the 25th percentile) to the third quartile (the 75th percentile). # The whiskers go from the minimum value to the maximum value unless the distance # from the minimum value to the first quartile is more than 1.5 times the IQR. # In that case the whisker extends out to the smallest value within 1.5 times # the IQR from the first quartile. A similar rule is used for values larger # than 1.5 times IQR from the third quartile. A special symbol shows the values, # called outliers, which are smaller or larger than the whiskers # boxplot(t(gr.oep.losses.order.20),names=round(20/c(20:1),2),xlab="Return Period",ylab="Losses",cex=1) title(main="Simulated EP Uncertainty",cex.main=1.7)

Illustration : Contribution to AAL by return period of simulated years (left). EP uncertainty (right) using 20 years of simulated data. Observed losses (if available) should be added. # This program is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or 21

# (at your option) any later version. # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # You should have received a copy of the GNU General Public License # along with this program. If not, see <http://www.gnu.org/licenses/>.

22