Download - Claims/Agency metrics
![Page 1: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/1.jpg)
1
Claims/Agency metrics
Greg Taylor
Taylor Fry Consulting Actuaries
University of Melbourne
University of New South Wales
Casualty Actuarial Society
Special Interest Seminar on Predictive Modeling
Boston, October 4-5 2006
![Page 2: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/2.jpg)
2
Overview
• Individual claim models• “Paids” models
• “Incurreds” models
• Numerical results
• Adaptive models
![Page 3: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/3.jpg)
3
Why individual claim models?
![Page 4: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/4.jpg)
4
Example problem
• Classical workers compensation cost centre allocation problem
• Claim numbers at the leaves of this tree may be small
Total claim cost
Cost centre 1
Cost centre 2
Cost centre m
. . .
… … …
![Page 5: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/5.jpg)
5
Measuring claims performance
• Consider measuring claims performance in a segment of a long tail portfolio
• Likely that adopted metric will require an estimate of the amount of losses incurred but as yet unpaid (loss reserve)• e.g. metric is expected ultimate losses per policy for a
specific underwriting period =
Paid to date + unpaid lossesNumber of policy-years of exposure
= average PTD per policy-year + average unpaid per policy-year
![Page 6: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/6.jpg)
6
Measuring claims performance in large portfolio segments• Let there be n policy-years of exposure and
ui = i-th amount unpaid
• Consider the ui to be random drawings from some distribution
• Average amount unpaid is
ūi = Σ ui /n = Σ {E[ui] + ui - E[ui]}/n
= E[ui] + Σ {ui - E[ui]}/n
E[ui] as n∞by the large of large numbers
d
![Page 7: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/7.jpg)
7
Measuring claims performance in large portfolio segments (cont’d)
ūi E[ui] as n∞• E[ui] = expected size of a randomly drawn claim• This will be the result produced by most
conventional actuarial methods, e.g.• Paid chain ladder• Even incurred chain ladder at early development
• While E[ui] may be a good approximation to ūi for large sample sizes, it may be very poor for small ones• Leading to a highly distorted cost allocation
d
![Page 8: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/8.jpg)
8
Measuring claims performance in small portfolio segments
• Effective estimation of small sample average claim cost must somehow take account of the properties of the individual claims
![Page 9: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/9.jpg)
9
There is a need to change from this…
Data Fitted
Forecast
Model Forecast
Conventional actuarial analysis of loss experience
• Call such models “aggregate models”
![Page 10: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/10.jpg)
10
…to this
Forecast
Claim 1
Claim 2
Claim 3
Claim n
:
:
:
Claim 1
Claim 2
Claim 3
Claim n
:
:
:
Model
Claim 1
Claim 2
Claim 3
Claim n
:
:
:
Special case of individual claim reserving – statistical case estimation
![Page 11: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/11.jpg)
11
Individual claim models
![Page 12: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/12.jpg)
12
Form of such a model
Claim 1
Claim 2
Claim 3
Claim n
:
:
:
Claim 1
Claim 2
Claim 3
Claim n
:
:
:
Claim 1
Claim 2
Claim 3
Claim n
:
:
:
Model
Y=f(β)+ε
Forecast
g( )
![Page 13: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/13.jpg)
13
Form of such a model
Claim 1
Claim 2
Claim 3
Claim n
:
:
:
Claim 1
Claim 2
Claim 3
Claim n
:
:
:
Claim 1
Claim 2
Claim 3
Claim n
:
:
:
Model
Y=f(β)+ε
Yi = f(Xi; β) + εi
Yi = size of i-th completed claim
Xi = vector of attributes (covariates) of i-th claim
β = vector of parameters that apply to all claims
εi = vector of centred stochastic error terms
Forecast
g( )
![Page 14: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/14.jpg)
14
Form of individual claim model
Yi = f(Xi; β) + εi
• Convenient practical form is
Yi = h-1(XiT β) + εi [GLM form]
h = link function Error distribution from exponential dispersion family
Linear predictor = linear function of the parameter vector
![Page 15: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/15.jpg)
15
Form of individual claim model – more specifically• How might one create an individual claim model of the
“paids” type?• Aggregate paids model usually takes the form
Yjk = f(j,k; β) + εjk
for
j = accident period
k= development period
• Compare with
Yi = f(Xi; β) + εi
Not always formulated
![Page 16: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/16.jpg)
16
Form of “paids” individual claim model
• Possible to mimic aggregate model by defining individual model as just
Yi = h-1(ji,ki; β) + εi
![Page 17: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/17.jpg)
17
Form of “paids” individual claim model
• Possible to mimic aggregate model by defining individual model as just
Yi = h-1(ji,ki; β) + εi
• But often possible to improve on this, e.g.• Replace development period j with operational
time ti (proportion of accident period’s incurred claims completed) at completion of i-th claim
• Example
Yi = exp [β0+β1ti+β2max(0,ti-0.8)] + εi
![Page 18: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/18.jpg)
18
Example of “paids” individual claim model
Yi = exp [β0+β1ti+β2max(0,ti-0.8)] + εi
E[Yi] = exp [β0+β1ti+β2max(0,ti-0.8)]Linear predictor of expected claim size as a function of operational
time
0
2
4
6
8
10
12
0
0.04
0.08
0.12
0.16 0.
2
0.24
0.28
0.32
0.36 0.
4
0.44
0.48
0.52
0.56 0.
6
0.64
0.68
0.72
0.76 0.
8
0.84
0.88
0.92
0.96
1
Operational time
Expected claim size as a function of operational time
$0
$10,000
$20,000
$30,000
$40,000
$50,000
$60,000
$70,000
0
0.04
0.08
0.12
0.16 0.
2
0.24
0.28
0.32
0.36 0.
4
0.44
0.48
0.52
0.56 0.
6
0.64
0.68
0.72
0.76 0.
8
0.84
0.88
0.92
0.96
1
Operational time
![Page 19: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/19.jpg)
19
Example of “paids” individual claim model (cont’d)
Yi = exp [β0+β1ti+β2max(0,ti-0.8)] + εi
• Include superimposed inflation• Let q=j+k=calendar period of claim
completion• Extend model
Yi = exp [β0+β1ti+β2max(0,ti-0.8)+β3qi] + εi
• Superimposed inflation at rate exp[β3] per period
![Page 20: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/20.jpg)
20
Example of “paids” individual claim model (cont’d)
Yi = exp [β0+β1ti+β2max(0,ti-0.8)+β3qi] + εi
• We might wish to model superimposed inflation as beginning at period q=q0
Yi = exp [β0+β1ti+β2max(0,ti-0.8)+β3max(0,qi-q0)] + εi
![Page 21: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/21.jpg)
21
Example of “paids” individual claim model (cont’d)
Yi = exp [β0+β1ti+β2max(0,ti-0.8)+β3qi] + εi
• We might wish to model superimposed inflation as beginning at period q=q0
Yi = exp [β0+β1ti+β2max(0,ti-0.8)+β3max(0,qi-q0)] + εi
• …and we might wish to model superimposed inflation with a rate that decreases with increasing operational time
Yi = exp [β0+β1ti+β2max(0,ti-0.8)+(β3-β4ti) max(0,qi-q0)] + εi
etc etc
![Page 22: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/22.jpg)
22
Example of “paids” individual claim model (cont’d) • “Paids” estimate of
loss reserve scaled to baseline $1,000M
• Prediction CoV = 5.3%
• Mack (incurreds) estimate is $887M with CoV = 10.5%
• Mack estimate produces negative reserves for the old years of origin
• “Paids” chain ladder fails completely
0
50
100
150
200
250
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Accident Year
Out
stan
ding
Lia
bilit
y ($
M)
Mack
Mack -1 s.d
Mack +1 s.d
Time Paids
Time Paids -1 s.d
Time Paids +1 s.d
![Page 23: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/23.jpg)
23
Example of “paids” individual claim model (cont’d)
• This model is very economical
• Contains only 9 parameters to represent many thousands of claims
Name Value Standard Error Standard Error (%)Mean 9.755 0.18697 1.9
l_Qtrly optime(^1) 0.0086 0.00167 19.5
l_Qtrly optime<12(^2) -0.0158 0.00049 3.1
l_Finalisation year<1981(^1) 0.1917 0.02029 10.6
l_Finalisation year1981-87(^1) 0.0795 0.00844 10.6
l_Finalisation year1987-90(^1) 0.0244 0.01281 52.4
l_Finalisation year>1990(^1) 0.0386 0.00379 9.8
l-Qtrly optime<70(^1) 0.0034 0.00204 60.3
l_Finalisation year>1990(^1).l-Qtrly optime<70(^1) 0.0008 0.0001 12.7
![Page 24: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/24.jpg)
24
Further extension of “paids” individual claim model
Yi = exp [β0+β1ti+β2max(0,ti-0.8)+(β3-β4ti) max(0,qi-q0)] + εi
• May include claim characteristics other than time-related, e.g.• Nature of injury
• Claim severity (MAIS scale)
• Pre-injury earnings
Yi = exp [β0+β1ti+β2max(0,ti-0.8)+(β3-β4ti) max(0,qi-q0) +
more terms] + εi
![Page 25: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/25.jpg)
25
Example of “incurreds” individual claim model
• Similar to “paids” model
• Basic set-up is still
Yi = h-1(ji,ki,qi,ti,other;β) + εi
• Example
Yi = exp(Ci,ji,ki,qi,ti,other;β) + εi
where Ci = current manual estimate of incurred cost of i-th claim
![Page 26: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/26.jpg)
26
Example of “incurreds” individual claim model (cont’d)• In fact, the model requires more structure than this because
of claims and estimates for nil cost
• Let (for an individual claim)• U = ultimate incurred (may = 0)
• C = current estimate (may = 0)
• X = other claim characteristics
Model of Prob[U=0|C,X]
Model of U|U>0,C=0,X
Model of U/C|U>0,C>0,X
Prob[U=0]
Prob[U>0]
If C=0
If C>0
![Page 27: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/27.jpg)
27
Example of “incurreds” individual claim model (cont’d) • “Paids” estimate of
loss reserve = $1,000M
• Prediction CoV = 5.3%
• “Incurreds” estimate of loss reserve = $1,040M
• Prediction CoV = 5.3%
0
0
0
1
10
100
1,000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Accident Year
Out
stan
ding
Lia
bilit
y ($
M)
Time Paids
Time Paids -1 s.d
Time Paids +1 s.d
Incurreds
Incurreds -1 s.d
Incurreds +1 s.d
![Page 28: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/28.jpg)
28
Adaptive reserving
![Page 29: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/29.jpg)
29
Static and dynamic models
• Return for a while to models based on aggregate (not individual claim) data
• Model form is still Y=f(β)+ε
• Example• j = accident quarter
• k = development quarter
• E[Yjk] = a kb exp(-ck) = exp [α+βln k - γk]
• (Hoerl curve for each accident period)
![Page 30: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/30.jpg)
30
Static and dynamic models (cont’d)
• ExampleE[Yjk] = a kb exp(-ck) = exp [α+βln k - γk]• Parameters are fixed• This is a static model
But parameters α, β, γ may vary (evolve) over time, e.g. with accident period
Then• E[Yjk] = exp [α(j)+β(j) ln k - γ(j) k]• This is a dynamic model, or adaptive model
![Page 31: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/31.jpg)
31
Illustrative example of evolving parameters
Separate curves represent different accident periods
0
10
20
30
40
50
60
70
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59
Development period
![Page 32: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/32.jpg)
32
Formal statement of dynamic model
• Suppose parameter evolution takes place over accident periods
• Y(j)=f(β(j)) +ε(j) [observation equation]
• β(j) = u(β(j-1)) + ξ(j) [system equation]
• Let (j|s) denote an estimate of β(j) based on only information up to time s
Some function Centred stochastic perturbation
![Page 33: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/33.jpg)
33
Adaptive reserving
(1|q)(2|q)
(q|q)
q-th diagonal
Forecast at valuation date
q
![Page 34: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/34.jpg)
34
Adaptive reserving (cont’d)
• Reserving by means of an adaptive model is adaptive reserving
• Parameter estimates evolve over time• Fitted model evolves over time• The objective here is “robotic reserving” in
which the fitted model changes to match changes in the data• This would replace the famous actuarial
“judgmental selection” of model
![Page 35: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/35.jpg)
35
Special case of dynamic model: DGLM
• Y(j)=f(β(j)) +ε(j) [observation equation]• β(j) = u(β(j-1)) + ξ(j) [system equation]• Special case:
• f(β(j)) = h-1(X(j) β(j)) for matrix X(j)• ε(j) has a distribution from the exponential dispersion family
• Each observation equation denotes a GLM• Link function h• Design matrix X(j)
• Whole system called a Dynamic Generalised Linear Model (DGLM)
![Page 36: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/36.jpg)
36
• Individual claim models can also be converted to adaptive form• Just subject parameters to evolutionary model
• We have experimented with this type of model and adaptive reserving• Moderately successful
Adaptive form of individual claim models
![Page 37: Claims/Agency metrics](https://reader035.vdocuments.site/reader035/viewer/2022062519/56814e4b550346895dbbd4c0/html5/thumbnails/37.jpg)
37
Conclusions
• Effective forecast of costs of small samples of claims requires individual claim models
• Such models condition the forecasts on much more information than aggregate models
• Even for large samples, individual claim models may yield considerably more efficient forecasts• Lower coefficient of variation
• This may save real money• Lower uncertainty implies lower capitalisation
• Adaptive forms of individual claim models may further improve the tracking of claims experience over time