SEVERE CONVECTIVE STORM MODELING Meghan Purdy Kay Cleary Associate Manager, Model Solutions Director, Regulatory Practice ©2013 Risk Management Solutions, Inc. A RECENT SURVEY What peril concerns you on a day-to-day basis? Has your company made changes to your severe weather ratemaking methodology in the last 3 years? ©2013 Risk Management Solutions, Inc. In your opinion, what is the biggest threat regarding climate change? 2 A RECENT SURVEY #1: SCS #2: Flood #3: Hurricane SCS, Flood, and Storm Surge/ Hurricane ~80% yes! ©2013 Risk Management Solutions, Inc. 3 RISK OVERVIEW Loss Historical Losses Accounts for 1/3 of all US peril AAL (~11 billion USD) Several events in last 15 years exceed $2 billion in loss – 3 events in 2011 – 1 so far in 2013 Risk Challenges High risk to aggregate covers, auto lines, and large single location risks Eats at profit, as most risk is retained Event frequency not well captured in statistical data SCS annual losses can be volatile/non-stable ©2013 Risk Management Solutions, Inc. 4 OUTLINE • • • ©2013 Risk Management Solutions, Inc. Intro to the RMS Severe Convective Storm model Applications and considerations Resilient risk management 5 SEVERE CONVECTIVE STORM MODELING FOUR PERILS OF SEVERE CONVECTIVE STORMS ©2013 Risk Management Solutions, Inc. • Hail – – – Most frequent of SCS perils Auto and Residential lines most at risk Smaller damage ratios, over large areas • Tornadoes – Rarest of the SCS perils – Highest damage ratios • Straight-line winds – Largest footprints of SCS perils – Treefall an issue for residential and auto • Lightning – Frequent, but least damaging – Losses to electrical equipment (power surge) 7 FRAMEWORK FOR SCS MODELING Generate Events ©2013 Risk Management Solutions, Inc. Assess Hazard Calculate Damage Quantify Financial Loss 8 EVENT GENERATION CHALLENGE: DEFINE THE PERIL ©2013 Risk Management Solutions, Inc. Image via foldedstory.com 10 CHALLENGE: DEFINE THE PERIL Tornadoes are #1 driver for loss of life – – 324 deaths in April 2011 outbreak Last death due to hail in US was 12 years ago; ~1,000 deaths due to tornadoes in same period Hail storms are #1 driver for insurance loss – – Aggregate loss: hail is dominant, 60% of all claims Tail loss: hail & tornado are ≈ 40% Annual Losses Large Event Losses Straight-line Wind Hail Tornado Straight-line Wind Tornado Hail Based on Claims Data ©2013 Risk Management Solutions, Inc. 11 CHALLENGE: BIASED HISTORICAL RECORDS Records and Observations (PCS) are limited to and biased by observation location and damage. Low Cat Models can provide physically-based frequency and severity distributions with complete coverage. Hazar High 12 Step 1 Step 2 Step 3 Step 4 SCS EVENT GENERATION Simulate stochastic years of atmospheric conditions • • Resample events from the North American Regional Reanalysis (NARR) – Reanalysis data from 19792005 Create “stochastic” years – 3-day blocks within 3 month periods – Over 27 years of data – Preserve seasonality – Preserve temporal and spatial correlations Stochastic Year 1 Stochastic Year 2 NARR Stochastic Day Day Year 12, Day 12 Day 1 Year 12, Day 13 Day 2 Year 12, Day 14 Day 3 NARR Stochastic Day Day Year 22, Day 12 Day 1 Year 22, Day 13 Day 2 Year 22, Day 34 Day 3 Year 03, Day 20 Year 03, Day 21 Year 03, Day 22 Year 05, Day 20 Year 05, Day 21 Year 05, Day 22 Day 4 Day 5 Day 6 Year 26, Day 33 Year 26, Day 34 Year 26, Day 35 Day 7 Day 8 Day 9 Year 14, Day 323 Day 365 Day 4 Day 5 Day 6 Year 27, Day 33 Year 27, Day 34 Year 27, Day 35 Day 7 Day 8 Day 9 Year 16, Day 311 Day 365 13 13 Step 1 Step 2 Step 3 Step 4 SCS EVENT GENERATION Create probability surface based on atmospheric conditions for an event • Combine NARR atmospheric conditions and historical observations – Create probability of specific perils occurring at a location – Climatology of risk Surface shear/CAPE triggers event initial location – CAPE (Convective Available Potential Energy, a measure of energy to feed storm development) – Wind Shear (to provide rotation for updrafts and tornadogenesis) – Size/intensity event as function shear/CAPE CAPE or shear atmospheric conditions # Historical Occurrences • CAPE or Shear 14 SCS EVENT GENERATION Perils & intensity based on probability surface and CAPE/Shear values • Contents of event modeled as function shear/CAPE independently • Individual peril intensity relationships – Derived from observations • Wind: anemometer network • Tornado: F-Scale • Hail: radar Straight Line Wind Intensity Hail Swath and Intensity Tornado Frequency & Intensity correlated Step 1 Step 2 Step 3 Step 4 Lightning Footprint 15 SCS EVENT GENERATION Apply peril footprints for event given probability and intensity information Step 1 Step 2 Step 3 Step 4 Hail Intensity1 / Intensity 2 Lightning Tornado Path • Some events will contain single, multiple, or all perils 16 SCS EVENT GENERATION: PUTTING IT ALL TOGETHER Shear ©2013 Risk Management Solutions, Inc. A hybrid model that unites statistics with numerical modeling • Numerical modeling provides thousands of years of large-scale, 3D meteorological “ingredients” for storms • Statistics are used to place tornado, hail, and straightline winds in each cell using probability distributions and historical data • Result is verified and calibrated against historical observations and damage surveys where appropriate CAPE 17 CHALLENGE: HIGH-FREQUENCY EVENTS State % AAL HF Alabama 9% Oklahoma 10% Texas 14% Louisiana 16% Wyoming 24% New York 28% Massachusetts 45% Nevada 77% Washington 82% ©2013 Risk Management Solutions, Inc. • • • High-frequency events can contribute over 50% of the annual AAL in some regions, particularly in the West Impractical to model as individual events SCS model’s solution: – Determine percentage of claims from high-frequency events, verify with CAPE as proxy for thunderstorms – 1 pseudo-event per state – Model as an annual occurrence (frequency = 1) for the aggregate contribution of high-frequency events to the location AAL 2011 IED, All Lines, All Subperils 18 CHALLENGE: HIGH-FREQUENCY EVENTS High-frequency event: Isolated t-storms/wind Low-frequency event: Major severe weather outbreak ©2013 Risk Management Solutions, Inc. Low-Frequency Events High-Frequency Events Storm Type Cat events Non-Cat events Examples Thunderstorms Straight-line winds Tornadoes Lightning Isolated Thunderstorms Downbursts Hailstorms Storm size Large-scale (1000s of sq mi or km) Small-scale (10s of sq mi or km) RiskLink Stochastic footprint? Yes No Regional Impact Dominant in Midwestern Plains Dominant in West 19 HAZARD HAIL Intensity 2 Hail Pad • Hailstorms - Intensity 1 Hail Pad Stochastic Hail Swath • Many hail swaths per day possible Calibrated with 50 years of observations Hail swaths often occur in clusters - Modeled at two intensity levels Intensity related to hail stone size and density Intensity distribution varies geographically Number of hail swaths, size, and intensity distribution dependent on storm size Footprint morphology calibrated on historical and radar data Ellipses fitted to the SPC points for the event of 3 May 1999, along with the WDT polygons from radar. ©2013 Risk Management Solutions, Inc. 21 STRAIGHT-LINE WINDS • Wind Surface Grid • • • • • MPH Ranging from microburst to derecho (1 mn/yr vs. 25 year) Derecho – widespread, long-lived convective windstorm Size: 3 miles to 100+ miles wide Duration: minutes to 24 hr Wind speeds: up to 100 mph gust Methods of reconstructing straight-line winds – Storm Prediction Center historical reports – Airport locations, mesonet stations, Global Summary of the Day – Examine roughness ©2013 Risk Management Solutions, Inc. 22 TORNADO • • • • Outbreak modeled by maximum F-intensity tornado Historical tornado reports are clustered into larger outbreaks (similar to hail) Intensity size distributions based on Rankine vortex model Adjusted with high-resolution damage surveys (from scientific literature, consultants) Vmax Tornado intensity based on Rankine vortex model. Goshen County WY: June 5th, 2009 ©2013 Risk Management Solutions, Inc. 23 LIGHTNING • • Losses from lightning strikes (non-fire) Two main damage modes: – – • • ©2013 Risk Management Solutions, Inc. Damage at point of entry (singe or burn marks) Electrical system (electronics that are plugged in) Typically low damage ratios Highly correlated with hail hazard so modeled on top 24 VULNERABILITY PERIL-SPECIFIC VULNERABILITY FUNCTIONS • • Photos from RMS (Matthew Nielsen) ©2013 Risk Management Solutions, Inc. • • • Distinct functions for Hail, Tornado, and Wind Hail kinetic energy – Key vulnerability components: • General roof shape (e.g. steep, low slope) • Roof cover (e.g. asphalt, shake, tile, built-up, single-ply) • Roof age (critical age ~10-15 years for most types) Tornado F-rating – Relates damage to approximate wind speed range Straight-line winds peak gust – Dominant range of wind speeds < 80 mph – Tree damage Use of claims data and consultants for calibration/validation 26 FUTURE MODEL UPDATES: RISKLINK ©2013 Risk Management Solutions, Inc. • Interim update of SCS model in January 2014 • Fundamentals of event generation module still strong • 2008-2012 taught us new lessons that we wish to integrate – Add information on tail events and EPs from 2008-2012 SCS seasons – Integrate new client data to further refine hazard and vulnerability 27 FUTURE MODEL UPDATES: RMS(ONE) EXPOSURE EVENT VULNERABILITY RATES Spring 2014: SCS translated for use on RMS(one) More powerful platform to make the model work for you: • • • Conduct sensitivity tests Leverage your own claims data and research Gain competitive advantage HAZARD PLA LOSS ©2013 Risk Management Solutions, Inc. 28 SCS APPLICATIONS & CONSIDERATIONS IMPLICATIONS AND APPLICATIONS • • • ©2013 Risk Management Solutions, Inc. Ratemaking (primary companies) – Statewide level – Territorial – Class Plans – Policy Terms Transfer of Risk (e.g., reinsurance) Concentration of Risk 30 HOW EVENTS ARE DEFINED SPC Risk Map *synoptic = large scale • Any vertically developed thunderstorm that produces damage due to hail, tornado, and/or a straight-line wind • Can occur in all states and provinces in the U.S. and Canada any time during the year • Peril model and catastrophe model • Event can be – Synoptic* system – Used in RiskLink to capture high-frequency losses atmospheric phenomenon ©2013 Risk Management Solutions, Inc. 31 EXPERIENCE DATA • Low frequency – PCS definition • • • • >=$25M industrywide, and >=$5M for any state Gross loss Lifetime of synoptic system – Company ID – ~$Ms • High frequency – Remainder – “follows” low freq – One “event” per year for each state – $10,000s to $100,000s ©2013 Risk Management Solutions, Inc. 32 HIGH FREQUENCY AND LOW FREQUENCY SCS LOSSES Contributes to AAL EP curve Discrete Events Low Freq Yes AEP / OEP Yes High Freq Yes Becomes meaningful when combined with lowfrequency losses Thousands of actual occurrences every year. ©2013 Risk Management Solutions, Inc. One event each year per state/province with varying hazard at more granular level. 33 RATES WITHIN A STATE OR REGION • • ©2013 Risk Management Solutions, Inc. Does geographic location within a state matter for SCS? Do you need to have territorial differentials? What about other characteristics? 34 AAL BY PRIMARY CHARACTERISTICS • Reference Structure: 200k structure, 150k contents, 40k ALE ($250 deductible) • Selected location in Midwest Scenario Construction Occupancy Yr Built # of Stories AAL CV 1 Unknown Unknown Unknown Unknown $82 32.7 2 Wood Unknown Unknown Unknown $107 27.0 3 Wood SFD Unknown Unknown $123 23.6 4 Wood SFD 1995 Unknown $113 25.0 5 Wood SFD 1995 2 $97 27.2 ©2013 Risk Management Solutions, Inc. 35 AAL BY PRIMARY CHARACTERISTICS Scenario Construction Occupancy Yr Built # of Stories AAL CV 6 Wood SFD 2005 2 $95 27.8 7 Wood SFD 1965 2 $107 25.5 Scenario Construction Occupancy Yr Built # of Stories AAL CV 6 Wood SFD 2005 2 $95 27.8 8 Wood SFD 2005 1 $115 24.7 ©2013 Risk Management Solutions, Inc. 36 PRIMARY CHARACTERISTICS: NUMBER OF STORIES 13% Damage Ratio Risk is primarily determined by the roof system covering and its value relative to the remainder of the structure – Brick veneer structure example – $100,000 per story replacement cost – $15,000 for roof 7% Damage Ratio 5% Damage Ratio 37 SECONDARY MODIFIERS Secondary modifiers are invoked only when sufficient primary characteristics are known: occupancy, construction class, year of construction, and building height Hail Tornado Straight-line Wind • Roof System Covering • Cladding Type • Roof Age • Mechanical and Electrical Systems • Foundation System • Roof Anchor • Wind Missiles • Tree Density • Cladding • Tree Density • Roof System Covering • Roof Sheathing Attachment 38 Return period of an F2 or greater tornado at a point in 1,000’s of years HOW DO YOU THINK ABOUT RISK AT A LOCATION BASIS? 100,000 50,000 20,000 10,000 5,000 2,000 • Location-level risk is fundamentally different for SCS than for other perils like hurricane • The RP of hurricane winds at a location is generally less than 100 years in high risk areas • The RP of an F2 at a location is measured in the THOUSANDS of years in high risk areas ©2013 Risk Management Solutions, Inc. Source: Meyer et al. 2002 39 DEDUCTIBLES • • Given that AAL is driven in large part by hail, damage ratios for SCS tend to be on the smaller side (5-10%) These types of loss ratios can be very sensitive to the deductible chosen when modeling SCS ©2013 Risk Management Solutions, Inc. Real-world case study: • Take a book of business for a particular state, and change the deductible from $250 to 1% of the limit • Determine the change to AAL and RP losses as a result Loss Metric / Return Period Change AAL -25% 5 -20% 10 -20% 50 -15% 100 -15% 250 -10% 500 -10% 43 RISK TRANSFER CONSIDERATIONS US TH AEP US HU AEP HU OEP US HUUSAEP US HU OEP 40% Exceedance Probability 40% Hurricane 45% US TH OEP US SCS AEP US SCS OEP 45% Tail of distribution Aggregate EP ~80% less than 72 hour duration 50% Severe Convective Storm 50% Exceedance Probability • • • 35% 30% 25% 20% 15% 10% 35% 30% 25% 20% 15% 10% 5% 5% 0% 0% 0 10 20 30 40 50 Gross Loss ©2013 Risk Management Solutions, Inc. 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Gross Loss 44 Looking at accumulation using a sample tornado footprint may be more helpful for understanding the amount of exposure at risk for ‘the big one’ EXPOSURE ACCUMULATION FOR TORNADOES Given the small footprint size for tornadoes, what are some ways to examine the worst-case scenarios? Develop a sample tornado ellipse using size parameters for long-lived/large tornadoes Place over hotspots of exposure and use simple damage ratios to calculate sample loss 45 MAY NEED ADDITIONAL EXPECTED $$ • Included - Tree fall - Debris removal - Power outage if there is direct damage to the location • Nonmodeled losses - Flood - Fire following - Power outage off premises unless there is direct damage to the location • Can model auto ©2013 Risk Management Solutions, Inc. 46 MORE INFORMATION • • • ©2013 Risk Management Solutions, Inc. RMS document in response to ASOP #38: Using Models Outside the Actuary’s Area of Expertise (Property and Casualty) Provides basic understanding of the model Non-proprietary – just ask 47 THE FUTURE: RESILIENT RISK MANAGEMENT RESILIENT RISK MANAGEMENT ©2013 Risk Management Solutions, Inc. Models aren’t perfect Resiliency in principal Resiliency in practice Catastrophe risk is characterized by deep uncertainty Understanding implied bets Diagnostic views and sensitivity tests Learning is ongoing Adapting quickly to new information Agile updates, post-event and interim views One size doesn’t always fit all Owning a view of risk Adjustments, alternatives, open platform 49 BENEFITS OF OWNING YOUR VIEW OF RISK More profitable and agile underwriting Improved capital allocation Take control of cat models Own View of Risk Reflect your unique portfolio ©2013 Risk Management Solutions, Inc. Leverage your experience and claims Stable view of risk over time Manage internal and external stakeholders 50 QUESTIONS? ©2013 Risk Management Solutions, Inc. • Kay.Cleary@rms.com | 850-386-5292 • Meghan.Purdy@rms.com | 510-608-3884 51 APPENDIX LA NIÑA AND ENHANCED TORNADO ACTIVITY El Niño The lack of strong jet stream energy generally leads to lower activity Sample analog years: 1992, 1998, 2003 La Niña A strong jet stream and warm moist air over the Southeast combine to create good conditions for storms Sample analog years: 1965, 1974, 2008, 2011 Maps courtesy of Climate Prediction Center 53 LOCATION OF EVENTS MATTERS Traditional ‘Tornado Alley’ is sparsely populated compared to areas of the South and East Tornadoes tend to happen earlier in the year in the South and East (Jan-April) and leads to higher fatality rates and increased losses ©2013 Risk Management Solutions, Inc. Source: Weather Channel 54 IS CLIMATE CHANGE A FACTOR IN SCS RISK? The effect of climate change on storms is difficult to discern for two reasons: • Historical record is not well resolved • Favorable SCS conditions are more tied to geography Storms to this point have not been proven to be more violent or more intense • EF4 and 5 tornado frequencies haven’t increased over time • EF0 and 1 tornadoes have seen increases, but most likely from historical underreporting than from any physical mechanism Your perception of the influence of climate change depends on how you trend historical data ©2013 Risk Management Solutions, Inc. 55 IS CLIMATE CHANGE A FACTOR IN SCS RISK? It is unclear how a warming climate will influence SCS behavior • Increase in warm, moist air should increase thunderstorms • Decrease in wind shear due to decrease in temperature gradient from equator to poles should lead to a decrease in hail and tornadoes • Strength and location of forcing mechanisms may lead to increases/drops in activity regionally Human Impacts • Outbreaks and severe weather peak months may shift to be earlier in the year • More people in harms way, as winter tornadoes tend to be more fatal ©2013 Risk Management Solutions, Inc. 56