The AIR Multiple Peril Crop Insurance Model for China

Transcription

1 The AIR Multiple Peril Crop Insurance Model for China Participation in China s Multiple Peril Crop Insurance (MPCI) program has dramatically increased since The growth in insurance penetration, together with complex and evolving policy conditions, means that relying on historical losses to estimate future losses is insufficient. The AIR MPCI Model for China provides a probabilistic approach and the most up-to-date view of the risk of losses arising from perils covered under China s crop insurance program.

2 A primary distinction between crop insurance and most other insurance lines is the correlation of losses across wide regions the result of large-scale adverse weather events. In mainland China, weather is to blame for 90% of all crop losses. The financial impact is significant. AIR estimates that a repeat of the 2000 drought would cost crop insurers more than 20 billion yuan today. Thus, in order to quantify the potential gains and losses to a crop insurance portfolio, it is critical to quantify the impact of weather. AIR s Comprehensive Approach to Modeling Crop Losses Offers Companies Multiple Views of Their Risk In 2011, AIR leveraged its considerable experience and success in modeling MPCI portfolios in the United States to develop a model for mainland China. Since then, we have updated the model several times to keep it up to date with the fast changing Chinese agriculture insurance market. The AIR MPCI Model for China captures the severity, frequency, and location of drought, flood, wind, and frost/freeze events nationwide, covering 90% of the weather-related crop losses. In addition, AIR also offers a probabilistic approach to model comprehensive insurance coverage of forests in China, accounting for fire, wind, precipitation, and pest/disease perils. AIR is dedicated to user flexibility, offering companies multiple views of their risk. Regular model updates ensure that analyses reflect the latest available weather, exposure, and policy condition information. In each update, AIR incorporates new crop area data from the China Statistical Yearbook and new policy condition information from the industry. Clients can generate custom results by utilizing peril and line-of-business filters, adjusting premium rates and sum insured per mu, and estimating model output from user-input deductibles (straight and franchise). In addition, the model provides a 10,000-year stochastic event catalog and a historical event set from 1981 to Premiums Claims CNY Billions TOTAL NATIONAL AGRICULTURAL PREMIUMS AND CLAIMS SINCE 1998 The growth in both premiums and claims arising from China s crop insurance program highlights the importance of modeling the impact of adverse weather events on China s cropland. (Sources: National Bureau of Statistics of China, China Insurance Regulatory Commission) 2

3 Accounting for the Major Crops Covered by Insurance The AIR MPCI Model for China estimates damage to the country s major crops: corn, cotton, rapeseed, rice, soybean, and wheat. Depending on the stage in the growing season at which an adverse weather event occurs, crop damage and the losses that result can be more, or less, severe. At the start of a growing season, for example, farmers have invested limited time and money into their crops; thus, if a weather event occurs, the potential losses are limited. As the season continues, however, the potential for losses grows. Insurance policy conditions in China cover crop production costs up to the time of damage, thus payouts are directly correlated to the crop s stage of development. Policy conditions also vary depending on the crop type, the peril, and the province. The AIR MPCI Model for China was built to accommodate these complex policy conditions. Leveraging Local Data for a Complete View of Agricultural Risk amid a Highly Variable Climate Mainland China has multiple climatic zones, ranging from subtropical to subarctic, and is subject to a wide range of weather events. The AIR MPCI Model for China captures the effects of frost, wind, drought, flood (where flooding can be due to excessive local precipitation, the runoff from more remote precipitation, and/or snowmelt), pest/disease (forests only), and fire (forests only). The model also captures the geographic variation in weather effects (for example, drought is more common in the arid north and west) and their different impacts (for example, wilt in the case of drought, or crop rot in the case of flood). To create a comprehensive 10,000-year catalog of simulated weather events, AIR scientists collected data on historical droughts, floods, typhoons, severe wind and frost events, forest fires, and pest/disease damage from various agencies, such as the U.S. National Oceanic and Atmospheric Administration and the National Bureau of Statistics of China. Daily temperature, precipitation, and wind speed data at high spatial resolution are also analyzed, and all this information is coupled with data on soils, terrain/elevation, and the spatial distribution of land cover types. The AWI Captures a Season of Drought and Flood Effects The AIR MPCI Model for China employs AIR s sophisticated Agricultural Weather Index TM (AWI TM ) to accurately capture effects of the severity, frequency, and location of adverse weather events, while also correctly preserving the influence of the timing of events during the season. The AWI assesses a region s water availability compared with each crop s Planted hectares reveal that only a fraction of China s vast expanse is arable. Water in the main growing areas of the north is scarce, while in the south, flooding is the dominant risk. Across China, farmers are extremely vulnerable to weather variability. Planted Hectares in each 50 km x 50 km grid cell (all modeled crops combined) 25,000-55,000 55,000-90,000 90, , , , , ,000 3

4 Elevation Available Water Capacity Land Use/Land Cover AWI TM Daily Minimum Temperature Daily Precipitation Daily Maximum Temperature AIR s AWI is a measure of the impact of drought and flood effects on crops. Key inputs into the AWI include high-resolution temperature, precipitation, and soil data, along with crop-specific agronomic data. The effects of freezing temperatures and damaging winds on crops are modeled separately. specific water needs and captures the season-long effects of drought and flood conditions on crops using time-series data on daily minimum and maximum temperature, daily accumulated precipitation, available water capacity of soil, crop-specific data (water requirement at each stage of crop development, planting dates, and resiliency to adverse weather conditions), land use/land cover, and terrain elevation. For temperature and precipitation, the model uses high-resolution gridded data of daily observations dating back to In the AIR model, drought intensity is quantified by the level of soil aridity above a threshold and its duration, while excess moisture intensity is quantified by the level of soil moisture above a threshold and its duration. The aridity and soil moisture are computed based on a water balance module that estimates the available water content at any point in a growing season. Information from the water balance module is combined with information from a module that accounts for flooding due to runoff from remote precipitation and snowmelt to determine the extent and duration of flood conditions. The final estimate of crop damage is made by using crop-specific, weather-based damage functions and accounting for the available moisture information. The resulting AWI index value describes the effects of the weather conditions experienced by a particular crop for the duration of the growing season. Creating New Crop Year Scenarios from Historical Data Including the effects of droughts, floods, high winds (including typhoons), and freezing temperatures, the AIR MPCI Model for China covers over 90% of the weatherrelated crop losses in China. Crop damage due to wind is determined from the combination of wind intensity (wind speed and duration) and crop-specific vulnerability. Frost intensity uses high-resolution temperature data to measure the extent that the minimum temperature falls beneath a threshold, taking into account crop-specific characteristics such as freezing points and growing season dates. 4

5 THE AIR MULTIPLE PERIL CROP INSURANCE MODEL FOR CHINA file_dir = "F:/Models/AgChina/data/stochasticCatalog/" files = dir(file_dir,".nc") filesv04 = files[grep("v04",files)] full_file_paths = paste(file_dir,filesv04,sep="") Actual file_dir = "F:/Models/AgChina/data/stochasticCatalog/" files = dir(file_dir,".nc") filesv04 = files[grep("v04",files)] full_file_paths = paste(file_dir,filesv04,sep="") nfiles = length(full_file_paths) nfiles = length(full_file_paths) cell_loc = read.csv("f:/models/agchina/data/staticdata/lonlatids csv") cell_loc = read.csv("f:/models/agchina/data/staticdata/lonlatids csv") loc_order_indx = order(cell_loc$cellid) cell_lon = cell_loc$lon[loc_order_indx] cell_lat = cell_loc$lat[loc_order_indx] cell_id = cell_loc$cellid[loc_order_indx] loc_order_indx = order(cell_loc$cellid) cell_lon = cell_loc$lon[loc_order_indx] cell_lat = cell_loc$lat[loc_order_indx] cell_id = cell_loc$cellid[loc_order_indx] cell_dim = dim(cell_loc)[1] event_indx = c(1,31,61,91,121,358) + 22 event_dim = length(event_indx) cell_dim = dim(cell_loc)[1] event_indx = c(1,31,61,91,121,358) + 22 event_dim = length(event_indx) week_indx = 35 week_indx = 35 cmi_map = c() #array(na,dim=c(event_dim,cell_dim)) cmi_hist_map = c() event = c() cell = c() cmi_map = c() #array(na,dim=c(event_dim,cell_dim)) cmi_hist_map = c() event = c() cell = c() for (i in 1:nfiles) { nc = open.ncdf(full_file_paths[i]) library(ncdf) cmi = get.var.ncdf(nc,"cmi",start=c(1,1,week_indx),count=c(-1,-1,1)) library(devemf) event = c(event,get.var.ncdf(nc,"event")) cell_tmp = get.var.ncdf(nc,"cell") rm(list=ls()); gc();gc();gc();gc() cmi_cells = 1:(length(cell_tmp)/2) cell = c(cell,cell_tmp[cmi_cells]) source("tools/gridplotter.r") cmi = cmi[,cmi_cells] source("f:/models/agchina/tools/utils/pcafuncs.r") for (i in 1:nfiles) { library(ncdf)nc = open.ncdf(full_file_paths[i]) cmi = get.var.ncdf(nc,"cmi",start=c(1,1,week_indx),count=c(-1,-1,1)) library(devemf) event = c(event,get.var.ncdf(nc,"event")) = get.var.ncdf(nc,"cell") rm(list=ls());cell_tmp gc();gc();gc();gc() cmi_cells = 1:(length(cell_tmp)/2) cell = c(cell,cell_tmp[cmi_cells]) source("tools/gridplotter.r") source("f:/models/agchina/tools/utils/pcafuncs.r") cmi = cmi[,cmi_cells] file_dir = "F:/Models/AgChina/data/stochasticCatalog/" # week = get.var.ncdf(nc,"week") # week = get.var.ncdf(nc,"week") file_dir = "F:/Models/AgChina/data/stochasticCatalog/" files = dir(file_dir,".nc") files = dir(file_dir,".nc") cmi_map = cbind(cmi_map,cmi[event_indx,]) cmi_map = cbind(cmi_map,cmi[event_indx,]) filesv04 = files[grep("v04",files)] filesv04 = files[grep("v04",files)] full_file_paths paste(file_dir,filesv04,sep="") pca=<list(center = get.var.ncdf(nc,"pcacenter"), pca <- list(center = get.var.ncdf(nc,"pcacenter"), full_file_paths = paste(file_dir,filesv04,sep="") scale = get.var.ncdf(nc,"pcascale"), scale = get.var.ncdf(nc,"pcascale"), nfiles = length(full_file_paths) rotation = get.var.ncdf(nc,"pcarot"), rotation = get.var.ncdf(nc,"pcarot"), nfiles = length(full_file_paths) x = get.var.ncdf(nc,"pcax")) x = get.var.ncdf(nc,"pcax")) cell_loc = read.csv("f:/models/agchina/data/staticdata/lonlatids csv") weekhist <- get.var.ncdf(nc,"weekdate") weekhist <- get.var.ncdf(nc,"weekdate") cell_loc = read.csv("f:/models/agchina/data/staticdata/lonlatids csv") cellhist <- get.var.ncdf(nc,"cellid") cellhist <- get.var.ncdf(nc,"cellid") flaghist <- get.var.ncdf(nc,"perilflag") loc_order_indx = order(cell_loc$cellid) flaghist <get.var.ncdf(nc,"perilflag") loc_order_indx = order(cell_loc$cellid) yearhist <- get.var.ncdf(nc,"year") cell_lon = cell_loc$lon[loc_order_indx] yearhist <- get.var.ncdf(nc,"year") cell_lon = cell_loc$lon[loc_order_indx] cell_lat = cell_loc$lat[loc_order_indx] cell_lat = cell_loc$lat[loc_order_indx] yea <- apply(weekhist,2,function(w) as.integer(format(strptime(w,format="%y%m%d"),format="%y"))) <- apply(weekhist,2,function(w) as.integer(format(strptime(w,format="%y%m%d"),format="%y")))cell_id = cell_loc$cellid[loc_order_indx] cell_idyea = cell_loc$cellid[loc_order_indx] mon <- apply(weekhist,2,function(w) as.integer(format(strptime(w,format="%y%m%d"),format="%m"))) mon <- apply(weekhist,2,function(w) as.integer(format(strptime(w,format="%y%m%d"),format="%m"))) day <- apply(weekhist,2,function(w) as.integer(format(strptime(w,format="%y%m%d"),format="%j"))) day=<-dim(cell_loc)[1] apply(weekhist,2,function(w) as.integer(format(strptime(w,format="%y%m%d"),format="%j"))) cell_dim = dim(cell_loc)[1] cell_dim event_indx = c(1,31,61,91,121,358) + 22 event_indx = c(1,31,61,91,121,358) + 22 <- which(flaghist==1) indxcmi <which(flaghist==1) event_dim =indxcmi length(event_indx) event_dim = length(event_indx) indxh <- which(flaghist==2) indxh <- which(flaghist==2) week_indx = 35 week_indx = 35 ######## reconstruct historical CMI ######## reconstruct historical CMI cmi_map = c() #array(na,dim=c(event_dim,cell_dim)) cmi_map = numcells c() #array(na,dim=c(event_dim,cell_dim)) <- dim(pca$x)[1] numcells <- dim(pca$x)[1] cmi_hist_map = c() cmi_hist_map = c() numweeks <- dim(pca$x)[2] numweeks <- dim(pca$x)[2] event numyears = c() event = c() numyears <- dim(pca$x)[3] <- dim(pca$x)[3] cell = cmihist c() cell = c() cmihist <- array(pcarecon(array(pca$x,c(numcells*numweeks,numyears)),pca$rotation,pca$scale,pca$center),c(numcells,numweeks,numyears)) <- array(pcarecon(array(pca$x,c(numcells*numweeks,numyears)),pca$rotation,pca$scale,pca$center),c(numcells,numweeks,numyears)) for (i in 1:nfiles) { mihist <- array(pcarecon(array(pca$x,c(numcells*numweeks,numyears)),pca$rotation,pca$scale,pca$center),c(numcells,numweeks,numyears)) close.ncdf(nc) open.ncdf(full_file_paths[i]) nc =cmi_hist_map = c(cmi_hist_map,cmihist[1:(numcells/2),week_indx,23]) cmi} = get.var.ncdf(nc,"cmi",start=c(1,1,week_indx),count=c(-1,-1,1)) event = c(event,get.var.ncdf(nc,"event")) cell_tmp = get.var.ncdf(nc,"cell") cell_order_indx = order(cell) cmi_cells = 1:(length(cell_tmp)/2) cell = cell[cell_order_indx] cellcmi_map = c(cell,cell_tmp[cmi_cells]) = cmi_map[,cell_order_indx] cmi_hist_map = cmi_hist_map[cell_order_indx] cmi = cmi[,cmi_cells] gridplotter(cell_lon,cell_lat,cmi_hist_map[],col=heat.colors(10)) for (i in 1:nfiles) { close.ncdf(nc) nc = open.ncdf(full_file_paths[i]) cmi_hist_map = c(cmi_hist_map,cmihist[1:(numcells/2),week_indx,23]) } cmi = get.var.ncdf(nc,"cmi",start=c(1,1,week_indx),count=c(-1,-1,1)) event = c(event,get.var.ncdf(nc,"event")) cell_order_indx = order(cell) cell_tmp = get.var.ncdf(nc,"cell") = cell[cell_order_indx] cmi_cellscell = 1:(length(cell_tmp)/2) cmi_map = cmi_map[,cell_order_indx] cell = c(cell,cell_tmp[cmi_cells]) cmi_hist_map = cmi_hist_map[cell_order_indx] cmi = cmi[,cmi_cells] gridplotter(cell_lon,cell_lat,cmi_hist_map[],col=heat.colors(10)) Crop growing conditions during historical floods and droughts are perturbed to produce the model s 10,000year catalog. Shown here are five realizations of perturbing the growing conditions during a historical flood event from May FORESTRY MODULE The AIR MPCI Model for China offers a probabilistic approach for estimating comprehensive insurance losses to China s forests, explicitly modeling fire, wind, precipitation, and pest/disease risks. Using satellitederived land use/land cover data and government reports, AIR has identified the locations of forests and their damage histories. The model incorporates detailed policy conditions for forestry by province. Historical flood, drought, wind, and frost events from 1979 to 2011 form the basis for generating the events that make up the model s catalog of simulated events. By changing, or perturbing, the growing conditions that were experienced during the historical years, a catalog representing a wide range of outcomes is produced each equally likely, but with potentially very different implications for insured losses. The AIR event generation process carefully maintains correlations in growing conditions in both space and time. These correlations are extremely important from a risk management perspective since they are the basis of any risk protection available from a well-diversified crop insurance portfolio. Damage Functions Estimate Total Damage by Region and Crop C HINA F ORES T C OVERAGE Hec tares per 50 km x 50 km grid c ell High : Low : 52 or below No Significant Forest The AIR model incorporates separate damage functions for drought, flood, wind, and frost that translate the intensity of the hazard affecting a crop or portfolio of crops into monetary loss. For drought and flood/excess precipitation conditions, intensity is defined in terms of output from the AWI. For wind and frost events, intensity is defined in terms of maximum daily wind speed and minimum daily temperature, respectively. 5

6 The damage functions in the AIR model vary by region due to different agricultural practices. For a given crop, weather damage is contingent on developmental stage. Corn plants, for example, are particularly susceptible to damage from droughts during the late vegetative and early reproductive stages. Cotton is highly susceptible to lack of oxygen (such as that caused by flooding) or too much or too little moisture during germination. Weather conditions favorable for one crop can be detrimental to another. A spring drought, for example, will not significantly impact winter wheat, which is close to maturity in the spring (having been planted in the fall). Indeed, dry conditions at harvest time can be beneficial to harvesting operations and physical access to fields. But a spring drought will damage spring corn because corn is in the stage of rapid growth at this time and needs sufficient water. Finally, with the exception of losses due to tropical cyclones, the effects of weather on crops are typically not immediate. Though a crop may suffer a setback during a dry spell, it may subsequently recover. It is the sum of all weather-related effects that determines the final damage to a crop as realized at harvest time for all explicitly modeled crops. Modeled Losses Are Validated Against Historical Losses Damaged areas are validated using historical data. The figure displays the results for Shanxi Province showing that the modeled drought damaged areas for the years 2009 to 2014 agree well with the historically observed damaged areas. The AIR MPCI Model for China calculates insured losses through the application of crop insurance policy conditions to the model s catalog of simulated events. Each policy type is unique and may be based on combinations of provinceaverage cost of production, perils, premium rates, sums insured, deductibles, and indemnity levels. To ensure the most reliable modeled loss estimates available, losses from the AIR MPCI Model for China are carefully validated against actual loss experience. Applications for Crop Insurers and Reinsurers MPCI programs are evaluated by applying each of the 10,000-year catalog outcomes and determining the insured retained loss. The probability distribution of total losses across the 10,000 simulated outcomes provides the measure of the risk of loss. This is expressed in terms of an exceedance probability distribution, characterized by the average (expected) annual gain/loss, and losses at selected exceedance probability levels, such as 10% (10-year return period), 5% (20-year return period), 1% (100-year return period), and 0.4% (250-year return period) exceedance probabilities. Crop insurance and reinsurance evaluations are performed in AIR s CATRADER software. Clients can generate custom results by utilizing peril and line-of-business filters, adjusting premium rates and sum insured per mu, and estimating model output from user-input deductibles. Crop insurers can evaluate alternative strategies in terms of expected profit versus potential risk. Reinsurers can price excess of loss and quota share programs and manage their entire portfolio. Fraction of Planted Area Shanxi Province Observed Drought Modeled Drought Year AIR modeled damaged areas for the years 2009 to 2014 compare well to the observed damaged areas. 6

7 The average annual loss (AAL) in the exceedance probability (EP) curve reflects historical loss ratios in China s crop insurance market. Tail events with large return periods, such as major droughts, flood, or wind events, can produce losses that far exceed the average. The industry EP curve shown in blue shows the size of tail events relative to the average annual loss for a countrywide exposure. When exposure is concentrated in an individual province, such as Shanxi shown in green, the size of losses in tail events can far outstrip the average loss. The EP curves shown here have been scaled to the same average annual loss. Percentage of AAL Industry EP Curve for China and Shanxi Province 400% 300% 200% 100% 0 AAL 5 yrs 10 yrs 20 yrs 50 yrs 100 yrs 250 yrs China Shanxi Model at a Glance Modeled Perils Model Domain Supported Geographic Resolution Vulnerability Module Covered Crops Drought, flood, wind, and frost damage for major crops; fire, wind, precipitation, and pest/ disease for forests Mainland China County and province Crop vulnerability varies by farming practice, peril, and crop developmental stage Corn, cotton, rapeseed, rice, soybean, and wheat are all explicitly modeled; all forest types modeled Historical Catalog Historical losses based on current exposure and coverage terms recast for the years 1981 through 2015 Model Options Modeled losses can be estimated for user-input deductibles by province, crop, and peril; model output can be adjusted for sum insured per mu and premium rate and distinguished by peril and line-of-business; forest losses can be modeled for fire only, or for all perils Model Highlights Calculates losses at 50 km resolution Provides a probabilistic catalog that takes into account the spatial and temporal correlations of crop losses Leverages the award-winning AWI to accurately isolate the impact of weather on crop damage at the county and province level Includes the first probabilistic approach for determining the likelihood of losses to China s forests Reflects individual insurance programs by province 7

8 ABOUT AIR WORLDWIDE AIR Worldwide (AIR) provides risk modeling solutions that make individuals, businesses, and society more resilient to extreme events. In 1987, AIR Worldwide founded the catastrophe modeling industry and today models the risk from natural catastrophes, terrorism, pandemics, casualty catastrophes, and cyber attacks, globally. Insurance, reinsurance, financial, corporate, and government clients rely on AIR s advanced science, software, and consulting services for catastrophe risk management, insurance-linked securities, site-specific engineering analyses, and agricultural risk management. AIR Worldwide, a Verisk (Nasdaq:VRSK) business, is headquartered in Boston with additional offices in North America, Europe, and Asia. For more information, please visit AIR Worldwide A Verisk Business