Commercial Property Size of Loss Distributions Glenn Meyers Insurance Services Office, Inc. Casualty Actuaries in Reinsurance June 15 , 2000 Boston, Massachusetts Outline • Data • Classification Strategy – Amount of Insurance – Occupancy Class • Mixed Exponential Model – “Credibility” Considerations • • • • Limited Classification Information Program Demonstration Goodness of Fit Tests Comparison with Ludwig Tables Separate Tables For • Commercial Property (AY 1991-95) • Sublines – BG1 (Fire and Lightning) – BG2 (Wind and Hail) – SCL (Special Causes of Loss) • Coverages – Building – Contents – Building + Contents – Building + Contents + Time Element Exposures • Reported separately for building and contents losses • Model is based on combined building and contents exposure – Even if time element losses are covered Classification Strategy • Amount of Insurance – Big buildings have larger losses – How much larger? • Occupancy Class Group – Determined by data availability • Not used – Construction Class – Protection Class Potential Credibility Problems • • • • Over 600,000 Occurrences 59 AOI Groupings 21 Occupancy Groups The groups could be “grouped” but: – Boundary discontinuities – We have another approach The Mixed Exponential Size of Loss Distribution 6 F( x) 1 w i e x/ i i1 i’s vary by subline and coverage • wi’s vary by AOI and occupancy group in addition to subline and coverage The Mixed Exponential Size of Loss Distribution 6 F( x) 1 w i e x/ i i1 i = mean of the ith exponential distribution • For higher i’s, a higher severity class will tend to have higher wi’s. The Fitting Strategy for each Subline/Coverage • Fit a single mixed exponential model to all occurrences • Choose the wi’s and i’s that maximize the likelihood of the model. • Toss out the wi’s but keep the i’s • The wi’s will be determined by the AOI and the occupancy group. Back to the Credibility Problem Raw Mixed Exponential Fits 1.2 1 0.8 W1 W2 0.6 W3 W4 0.4 W5 W6 0.2 0 1 10 100 1,000 10,000 100,000 -0.2 AOI (All Occupancy Groups Combined) 1,000,000 Back to the Credibility Problem Fitted Excess Severities 90000 Excess Severity over $10,000 80000 70000 60000 50000 40000 30000 20000 10000 0 1 10 100 1,000 10,000 100,000 -10000 AOI (All Occupancy Groups Combined) 1,000,000 Varying Wi’s by AOI Prior expectations • Larger AOIs will tend to have higher losses • In mixed exponential terminology, the AOI’s will tend to have higher wi’s for the higher i’s. • How do we make this happen? Solution • Let W1i’s be the weights for a given AOI. • Let W2i’s be the weights for a given higher AOI. • Given the W1i’s, determine the W2i’s as follows. Step 1 Choose 0 d11 1 W11 W21 = (1-d11) W11 W12 W22 = W12+d11W11 W13 W23 = W13 W16 W26 = W16 Shifting the weight from 1st exponential to the 2nd exponential increases the expected claim cost. Step 2 Choose 0 d12 1 W11 W21 = (1-d11) W11 W12 W22 = (1-d12) (W12+d11W11) W13 W23 = W13+d12 (W12+d11W11) W16 W26 = W16 Shifting the weight from 2nd exponential to the 3rd exponential increases the expected claim cost. Step 3 and 4 Similar Step 5 — Choose 0 d15 1 W11 W21 = (1-d11) W11 W12 W22 = (1-d12) (W12+d11W11) W13 W23 = (1-d13)(W13+d12(W12+d11W11)) 5 W16 W26 = 1 W2i i1 Shifting the weight from 5th exponential to the last exponential increases the expected claim cost. Several AOI Groups Choose W’s for lowest AOI Group W11 W12 W13 W14 W15 W16 Then choose d’s to Construct W’s for the 2nd AOI Group W11 W12 W13 W14 W15 W16 W21 W22 W23 W24 W25 W26 Then choose d’s to Construct W’s for the 3rd AOI Group W11 W12 W13 W14 W15 W16 W21 W22 W23 W24 W25 W26 W31 W32 W33 W34 W35 W36 Then choose d’s to Construct W’s for the 4th AOI Group W11 W12 W13 W14 W15 W16 W21 W22 W23 W24 W25 W26 W31 W32 W33 W34 W35 W36 W41 W42 W43 W44 W45 W46 Continue choosing d’s and constructing W’s until the end. W11 W12 W13 W14 W15 W16 W21 W22 W23 W24 W25 W26 W31 W32 W33 W34 W35 W36 W41 W42 W43 W44 W45 W46 Estimating W’s (for the 1st AOI Group) and d’s (for the rest) Let: • Fk(x) = CDF for kth AOI Group • (xh+1, xh) be the hth size of loss group • nhk = number of occurrences for h and k Then the log-likelihood of data is given by: Fk ( xh1) Fk ( xh )f nhk loga h k Estimating W’s (for the 1st AOI Group) and d’s (for the rest) • Choose W’s and d’s to maximize loglikelihood • 59 AOI Groups • 5 parameters per AOI Group • 295 parameters! Too many! Parameter Reduction • Fit W’s for AOI=1, and d’s for AOI=10, 100, 1,000, 10,000, 100,000 and 1,000,000. Note AOI coded in 1,000’s • The W’s are obtained by linear interpolation on log(AOI)’s • The interpolated W’s go into the loglikelihood function. • 35 parameters -- per occupancy group On to Occupancy Groups • Let W be a set of W’s that is used for all AOI amounts for an occupancy group. • Let X be the occurrence size data for all AOI amounts for an occupancy group. • Let L[X|W] be the likelihood of X given W i.e. the probability of X given W There’s No Theorem Like Bayes’ Theorem • Let kp n Wk k 1 be n parameter sets. • Then, by Bayes’ Theorem: k p l q Prl W X q LlX W q PrkW p k L X Wk Pr Wk n k 1 k k Bayesian Results Applied to an AOI and Occupancy Group • Let k wi be the ith weight that Wk assigns to the AOI/Occupancy Group. • Then the wi‘s for the AOI/Occupancy Group is: wi n l q k w i Pr Wk X k=1 What Does Bayes’ Theorem Give Us? • Before – A time consuming search for parameters – Credibility problems • If we can get suitable Wk’s we can reduce our search to n W’s. • If we can assign prior Pr{Wk}’s we can solve the credibility problem. Finding Suitable Wk’s • Select three Occupancy Class Group “Groups” • For each “Group” – Fit W’s varying by AOI – Find W’s corresponding to scale change • Scale factors from 0.500 to 2.000 by 0.025 • 183 Wk’s for each Subline/Coverage Neg. Average Log Likelihood Graph of Log-Likelihoods 2 1.98 1.96 1.94 1.92 1.9 1.88 1.86 1.84 1.82 1.8 0.000 High Medium Low 0.500 1.000 1.500 Scale Factor 2.000 2.500 Prior Probabilities kp • Set: Pr Wk 1 / n 1 / 183 • Final formula becomes: LlX W q k p l q Prl W X q LlX W q PrkW p LlX W q k L X Wk Pr Wk k n k 1 n k k k 1 • Can base update prior on Pr{Wk |X}. k The Classification Data Availability Problem • Focus on Reinsurance Treaties – Primary insurers report data in bulk to reinsurers – Property values in building size ranges – Some classification, state and deductible information • Reinsurers can use ISO demographic information to estimate effect of unreported data. Database Behind PSOLD 30,000+ records (for each coverage/line combination) containing: • Severity model parameters • Amount of insurance group – 59 AOI groups • Occupancy class group • State • Number of claims applicable to the record Constructing a Size of Loss Distribution Consistent with Available Data Using ISO Demographic Data • Select relevant data • Selection criteria can include: – Occupancy Class Group(s) – Amount of Insurance Range(s) – State(s) • Supply premium for each selection • Each state has different occupancy/class demographics Constructing a Size of Loss Distribution for a “Selection” • Record output - Layer Average Severity • Combine all records in selection: LASSelection = Wt Average(LASRecords) Use the record’s claim count as weights Constructing a Size of Loss Distribution for a “Selection” n w ij C j i j1 n Cj j1 Where: i = ith overall weight parameter wij = ith weight parameter for the jth record Cj = Claim weight for the jth record The Combined Size of Loss Distribution for Several “Selections” • Claim Weights for a “selection” are proportional to Premium Claim Severity LASCombined = Wt Average(LASSelection) Using the “selection” total claim weights • The definition of a “selection” is flexible The Combined Size of Loss Distribution for Several “Selections” • Calculate i’s for groups for which you have pure premium information. • Calculate the average severity for jth group 6 j ij i i1 The Combined Size of Loss Distribution for Several “Selections” • Calculate the group claim weights Pure Pr emium j j j • Calculate the weights for the treaty size of loss distribution n ij j i j1 n j j1 The Deductible Problem • The above discussion dealt with ground up coverage. • Most property insurance is sold with a deductible – A lot of different deductibles • We need a size of loss distribution net of deductibles Size of Loss Distributions Net of Deductibles Relative Frequency Relative Frequency • Remove losses below deductible • Subtract deductible from loss amount Loss Am ount Size of Loss Distributions Net of Deductibles • Combine over all deductibles LASCombined Post Deductible Equals Wt Average(LASSpecific Deductible) • Weights are the number of claims over each deductible. Size of Loss Distributions Net of Deductibles For an exponential distribution: Net severity Excess Pure Pr emium Excess Frequecy e d/ d/ e Ground up severity Need only adjust frequency -- i.e. wi’s Adjusting the wi’s • Dj jth deductible amount ij wi e 6 D j / i wi e D j / i i1 ded C • Wi j ij j Goodness of Fit - Summary • 16 Tables • Fits ranged from good to very good • Model LAS was not consistently over or under the empirical LAS for any table • Model unlimited average severity – Over empirical 8 times – Under empirical 8 times A Major Departure from Traditional Property Size of Loss Tabulations • Tabulate by dollars of insured value • Traditionally, property size of loss distributions have been tabulated by % of insured value. Fitted $ Average Severity against Insured Value Expected Loss at AOI 120,000 100,000 80,000 60,000 40,000 20,000 0 0 200,000 400,000 600,000 800,000 Amount of Insurance 1,000,000 1,200,000 Fitted Average Severity as % of Insured Value E[Loss] as a % of AOI 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% Blow up this area 0.00% 0 200,000 400,000 600,000 800,000 Amount of Insurance 1,000,000 1,200,000 Fitted Average Severity as % of Insured Value 60.00% E[Loss] as % of AOI 50.00% Eventually, assuming that loss distributions based on a percentage of AOI will produce layer costs that are too high. 40.00% 30.00% 20.00% 10.00% 0.00% 0 100 200 300 400 500 600 700 Amount of Insurance 800 900 1,000 PSOLD Demonstration • • • • No Information Size of Building Information Size + Class Information Size + Class + Location Information Comparison with Ludwig Tables • Tabulated by % of amount of insurance • Organized by occupancy class and amount of insurance – Broader AOI classes – Broader occupancy classes • Fewer occurrances • No model • A very good paper Comparison with Ludwig Tables • Ludwig — Exhibit 15 (all classes) • Matched insured value ranges • Obtained % of insured value distributions from PSOLD – assuming low end of range – assuming high end of range • Results on Spreadsheet What’s new for the next review? • Include data through 1998 • Fewer exclusions of loss information – Recall that we excluded claims if exposure and class information were missing. – Include claims if we trust the losses and use Bayesian techniques to spread losses to possible class and exposure groups. • Include HPR classes