Two Approaches to Calculating Correlated Reserve Indications Across Multiple Lines of Business...
-
Upload
randolf-berry -
Category
Documents
-
view
221 -
download
2
Transcript of Two Approaches to Calculating Correlated Reserve Indications Across Multiple Lines of Business...
Two Approaches to Calculating Correlated Reserve Indications Across Multiple Lines of Business
Gerald KirschnerClassic Solutions
Casualty Loss Reserve Seminar
September 2002
Presentation Structure Background – simulating reserves
stochastically Correlation – a general discussion Case study
Description Results using correlation matrix
approach Results using bootstrap approach
Conclusion
Simulating reserves stochastically using a chain-ladder method
Begin with a traditional loss triangle
Calculate link ratios
Calculate mean and standard deviation of the link ratios
Acc.Year 12 24 36 48
1 1,000 1,500 1,750 2,0002 1,200 2,000 2,3003 1,800 2,5004 2,100
Development Age
Acc.Year 12 - 24 24 - 36 36 - 48
1 1.500 1.167 1.1432 1.667 1.1503 1.389
Mean 1.500 1.157 1.143Std. Deviation 0.1179 0.0082 0
Link Ratios
Simulating reserves stochastically using a chain-ladder method Think of the observed link ratios for
each development period as coming from an underlying distribution with mean and standard deviation as calculated on the previous slide
Make an assumption about the shape of the underlying distribution – easiest assumptions are Lognormal or Normal
Simulating reserves stochastically using a chain-ladder method
For each link ratio that is needed to square the original triangle, pull a value at random from the distribution described by
1. Shape assumption (i.e. Lognormal or Normal)
2. Mean3. Standard deviation
Acc.Year 12 - 24 24 - 36 36 - 48
1 1.500 1.167 1.1432 1.667 1.1503 1.389
Mean 1.500 1.157 1.143Std. Deviation 0.1179 0.0082 0
Acc.Year 12 - 24 24 - 36 36 - 48
1 1.500 1.167 1.143
2 1.667 1.150 1.143
3 1.389 1.163 1.1434 1.419 1.145 1.143
Link Ratios
Link Ratios
Simulating reserves stochastically using a chain-ladder method
Simulated values are shown in red
Lognormal Distribution, mean 1.5, standard deviation 0.1179
0.0%
5.0%
10.0%
15.0%
20.0%
25.0%
30.0%
1.168 1.249 1.330 1.411 1.492 1.573 1.654 1.735 1.816 1.897
% o
f T
ota
l O
bs
erv
ati
on
s
Random draw
Simulating reserves stochastically using a chain-ladder method Square the triangle using the
simulated link ratios to project one possible set of ultimate accident year values. Sum the accident year results to get a total reserve indication.
Repeat 1,000 or 5,000 or 10,000 times.
Result is a range of outcomes.
Simulating reserves stochastically using other methods Numerous options for enhancing this
basic approach Logarithmic transformation of link ratios
before fitting Inclusion of a parameter risk adjustment Use of incremental data instead of
cumulative Extend payout via tail factor
extrapolation
Simulating reserves stochastically via bootstrapping
Bootstrapping is a different way of arriving at the same place
Bootstrapping does not care about the underlying distribution – instead bootstrapping assumes that the historical observations contain sufficient variability in their own right to help us predict the future
Actual Cumulative Historical Data
Acc.Year 12 24 36 48
1 1,000 1,500 1,750 2,0002 1,200 2,000 2,3003 1,800 2,5004 2,100
Ave Link Ratio 1.500 1.157 1.143
Recast Cumulative Historical Data
Acc.Year 12 24 36 48
1 1,008 1,512 1,750 2,0002 1,325 1,988 2,3003 1,667 2,5004 2,100
Development Age
Development Age
Simulating reserves stochastically via bootstrapping
1. Keep current diagonal intact
2. Apply average link ratios to “back-cast” a series of fitted historical payments
Ex: 1,988 =
2,300
Simulating reserves stochastically via bootstrapping
Actual Incremental Historical Data
Acc.Year 12 24 36 48
1 1,000 500 250 2502 1,200 800 3003 1,800 7004 2,100
Recast Cumulative Historical Data
Acc.Year 12 24 36 48
1 1,008 504 238 2502 1,325 663 3123 1,667 8334 2,100
Development Age
Development Age
3. Convert both actual and fitted triangles to incrementals
4. Look at difference between fitted and actual payments to develop a set of ResidualsResiduals
Acc.Year 12 24 36 48
1 (0.259) (0.183) 0.801 0.0002 (3.437) 5.340 (0.699)3 3.266 (4.619)4 0.000
Development Age
Simulating reserves stochastically via bootstrapping
5. Create a “false history” by making random draws, with replacement, from the triangle of residuals. Combine the random draws with the recast historical data to come up with the “false history”.
Random Draw from Residuals
Acc.Year 12 24 36 48
1 0.801 (0.183) 3.266 0.8012 5.340 (4.619) (0.259)3 (0.699) (3.437)4 5.340
False History
Acc.Year 12 24 36 48
1 1,034 500 288 2632 1,519 544 3083 1,638 7344 2,345
Development Age
Development Age
Simulating reserves stochastically via bootstrapping
6. Calculate link ratios from the data in the cumulated false history triangle
7. Use the link ratios to square the false history data triangle
8. Several additional steps are described in the paper…
9. Repeat process N times to get N different reserve indications.
Cumulated False History
Acc.Year 12 24 36 48
1 1,034 1,534 1,822 2,0842 1,519 2,063 2,3713 1,638 2,3724 2,345
Ave Link Ratio 1.424 1.166 1.144
Squaring of the Cumulated False History
Acc.Year 12 24 36 48
1 1,034 1,534 1,822 2,0842 1,519 2,063 2,371 2,7133 1,638 1,638 1,909 2,1854 2,345 3,339 3,892 4,454
Development Age
Development Age
Pros / Cons of each methodChain-ladder Pros More flexible - not
limited by observed data
Chain-ladder Cons
More assumptions Potential
problems with negative values
Bootstrap Pros Do not need to
make assumptions about underlying distribution
Bootstrap Cons Variability limited
to that which is in the historical data
Correlation Correlation vs. causality
Correlation is a a way of measuring the “strength of relationship” between two sets of numbers.
Causality is the relation between a cause (something that brings about a result) and its effect.
Can have correlation between two things without causality – both could be influenced by an unknown third item.
Effects of Correlation Suppose we have two lines, A & B, whose
reserve indications exhibit correlation Strength of the correlation is irrelevant if
we only care about the mean reserve indication for A + B:
mean (A + B) = mean (A) + mean (B) Strength of correlation matters when we
look towards the ends of the distribution of (A+B).
Effects of Correlation: Example 1
2 lines of business, N (100,25) 75th percentile of A+B at different levels
of correlation between A and B:
Correlation Values at 75th percentile
Ratio of Values at 75th percentile
0.00 223.8 0.0%
0.25 226.7 1.3%
0.50 229.2 2.4%
0.75 231.5 3.4%
1.00 233.7 4.4%
Effects of Correlation: Example 2
Same idea, but increase variability of distributions for lines A and B:
Standard Deviation Value 25 50 100 200 Value for 0.00 correlation
at the 75th percentile 223.8 247.7 295.4 390.8
Correlation Ratio of values at 75th percentile 0.25 1.3% 2.3% 3.8% 5.8% 0.50 2.4% 4.3% 7.3% 11.0% 0.75 3.4% 6.2% 10.4% 15.8% 1.00 4.4% 8.0% 13.4% 20.2%
Correlation methodologies Method 1: relies on the user to specify a
correlation matrix that describes the relative strength of relationship between the lines of business by analyzed. Will use rank correlation technique to develop a correlated reserve indication.
Method 2: uses the bootstrap process to maintain any correlations that might be implicit in the historical data. No other information is needed to develop the correlated reserve indication.
Rank correlation examplePerfectly Correlated
A B A B A B 5 4 1 1 5 3 4 1 2 2 4 2 2 5 3 3 2 5 1 2 4 4 1 1
Index A B 3 3 5 5 3 41 155 1542 138 1253 164 100 A B A+B A B A+B A B A+B4 122 198 107 198 305 155 154 309 107 100 2075 107 128 122 154 276 138 125 263 122 125 247
138 128 266 164 100 264 138 128 266155 125 280 122 198 320 155 154 309164 100 264 107 128 235 164 198 362
Low 264 Low 235 Low 207High 305 High 320 High 362
Rank to Use
Resulting Joint Dist.
Range of Joint Dist.
No CorrelationRank to Use
Resulting Joint Dist.
Range of Joint Dist.
Rank to Use
Resulting Joint Dist.
Range of Joint Dist.
Perfect Inverse Correlation
Method 1 approach Model user must determine (through
other means) the relative relationships between the lines of business being modeled
Information is entered into a correlation matrix
Uncorrelated reserve indications generated for each line and sorted from low to high
Method 1 approach continued Model creates a Normal distribution for
each line with mean = average reserve indication for each class, standard deviation = standard deviation of reserve indications for each class Normal distributions are correlated using the user-entered correlation matrix
Pull values from correlated normal distribution – drawing N correlated values, where N = # simulated reserve indications
Method 1 approach continued Use relative positioning of the
correlated Normal draws as the basis for pulling values from the sorted table of uncorrelated reserve indications to create correlated reserve indications across the lines of business
Example of Method 1 approach
2nd value from Class A, 3rd value from Class B
Iteration Class A Class B1 100 10002 200 20003 300 30004 400 40005 500 5000
Uncorrelated reservesSorted from low to high
1
Iteration Class A Class B1 3 52 5 23 1 14 4 45 2 3
Ranking of drawsfrom Correlated Normal
2 3
Iteration Class A Class B12345
Correlated reserves
Iteration Class A Class B1 300 50002 500 20003 100 10004 400 40005 200 3000
Correlated reserves
Iteration Class A Class B12345 200 3000
Correlated reserves
Bootstrap Correlation Methodology
1 2 3 4 1 2 3 4
1 (1A,1A) (1A,2A) (1A,3A) (1A,4A) 1 (1B,1B) (1B,2B) (1B,3B) (1B,4B)
2 (2A,1A) (2A,2A) (2A,3A) 2 (2B,1B) (2B,2B) (2B,3B)
3 (3A,1A) (3A,2A) 3 (3B,1B) (3B,2B)
4 (4A,1A) 4 (4B,1B)
1 2 3 4 1 2 3 4
1 (2A,1A) (3A,2A) (1A,3A) (3A,1A) 1 (2B,1B) (3B,2B) (1B,3B) (3B,1B)
2 (2A,2A) (1A,2A) (2A,3A) 2 (2B,2B) (1B,2B) (2B,3B)
3 (3A,1A) (1A,1A) 3 (3B,1B) (1B,1B)
4 (1A,1A) 4 (1B,1B)
Development Year Development Year
Correlated Bootstrapping - Reshuffling of variability parameters in Triangle B
Variability Parameters Calculated from Original Data
Triangle A Triangle BDevelopment Year Development Year
AY
AY
AY
AY
Calculated Variability Parameters
Randomly Selected Variability Parameters to be used in the creation of one possible pseudo-history
Pros / Cons of each approachCorrelation Matrix Pros More flexible - not
limited by observed data
Correlation Matrix Cons
Requires modeler to do additional work to quantify the correlations between lines
Bootstrap Correlation Pros
Do not need to make assumptions about underlying correlations
Bootstrap Cons Results reflect only
those correlations that were in the historical data
Case Study Three lines of business All produce approximately the same
mean reserve indication, but with different levels of volatility around the mean
Run a 5,000 iteration simulation exercise for each line
Examine the results for the aggregated reserve indication at different percentiles of the aggregate distribution
Rank correlation results
2,000,000
3,000,000
4,000,000
5,000,000
6,000,000
7,000,000
8,000,000
9,000,000
10,000,000
11,000,000
1%ile
5%ile
10%ile
20%ile
30%ile
40%ile
50%ile
60%ile
70%ile
80%ile
90%ile
95%ile
99%ile
Do
llars
(00
0 o
mit
ted
)
0% corr
25% corr
50% corr
75% corr
100% corr
Mean Value
75th PercentilePercentChange
from Zero Percentile
0% corr. 4,640,039 n/a25% corr. 4,697,602 1.2%50% corr. 4,739,459 2.1%75% corr. 4,794,767 3.3%100% corr. 4,836,166 4.2%
Estimated 75th Percentiles
Add Bootstrap results
2,000,000
3,000,000
4,000,000
5,000,000
6,000,000
7,000,000
8,000,000
9,000,000
10,000,000
11,000,000
1%ile
5%ile
10%ile
20%ile
30%ile
40%ile
50%ile
60%ile
70%ile
80%ile
90%ile
95%ile
99%ile
Do
llars
(00
0 o
mit
ted
)
0% corr
25% corr
50% corr
75% corr
100% corr
Bootstrap
Mean Value
75th PercentilePercentChange
from Zero Percentile
0% corr. 4,640,039 n/a25% corr. 4,697,602 1.2%50% corr. 4,739,459 2.1%75% corr. 4,794,767 3.3%100% corr. 4,836,166 4.2%
Bootstrap 4,755,952 2.5%
Estimated 75th Percentiles
Case Study Conclusions Mean aggregated reserve = 4.33B Reserves at the 75th percentile
range from 4.64B to 4.84B Bootstrap tells us that there does
appear to be some level of correlation in underlying data
General Conclusions To calculate an aggregate reserve
distribution, must understand and be able to quantify the dependencies between underlying lines of business
Correlation is probably not an important issue for lines of business with non-volatile reserve ranges, but can be important for ones with volatile reserves, especially as one moves further towards a tail of the aggregate distribution