# Heteroskedasticity in Loss Reserving CASE Fall 2012

```Heteroskedasticity in Loss
Reserving
CASE Fall 2012
What’s the proper link ratio?
Cumulative
Paid @12
Months
21,898
22,549
23,881
25,897
23,486
27,029
25,845
25,415
32,804
Cumulative
Paid @24 12-24 month
Months
LDF
56,339
2.5728
59,459
2.6369
60,315
2.5256
71,409
2.7574
59,165
2.5192
60,778
2.2486
60,543
2.3425
54,791
2.1559
59,141
1.8029
60,000
50,000
Subsequent Incremental
40,000
30,000
20,000
10,000
0
0
5,000
10,000
15,000
20,000
25,000
Prior Cumulative Losses
30,000
35,000
40,000
45,000
60,000
50,000
Subsequent Incremental
40,000
30,000
20,000
10,000
0
0
5,000
10,000
15,000
20,000
25,000
Prior Cumulative Losses
30,000
35,000
40,000
45,000
60000
50000
Incremental
40000
30000
20000
10000
0
0
5000
10000
15000
20000
25000
Prior Cumulative Losses
30000
35000
40000
45000
And the boring version
Method
Simple average
Weighted average
Unweighted least squares
LDF
2.3958
2.3686
2.3375
All of those estimators are unbiased.
Which one is efficient?
You’re not weighting link ratios.
You’re making an assumption
about the variance of the
observed data.
60000
50000
Incremental
40000
30000
20000
10000
0
0
5000
10000
15000
20000
25000
Prior Cumulative Losses
30000
35000
40000
45000
So how do we articulate our
variance assumptions?
A triangle is really a matrix
• A variable of interest (paid losses, for example) presumed to have some statistical
relationship to one or more other variables in the matrix.
• The strength of that relationship may be established by creating models which relate
two variables.
• A third variable is introduced by categorizing the predictors.
• Development lag is generally used as the category.
The response variable
will generally be
incremental paid or
incurred losses.
 y1   x11  x1 p   1   e1 
   
   




   
  
 yn   xn1  xnp    p  e p 
The design matrix may be either
prior period cumulative losses,
earned premium or some other
variable. Columns are
differentiated by category.
We assume that
error terms are
homoskedastic
and normally
distributed
Calibrated model
factors are analogous
to age-to-age
development factors.
 y1   x11  x1 p   1   e1 
   
   










 
 yn   xn1  xnp    p  e p 
21,898
22,549
23,881
25,897
23,486
27,029
25,845
25,415
32,804
40,409
56,339
59,459
60,315
71,409
59,165
60,778
60,543
54,791
59,141
78,457 92,820 101,261
82,918 98,768 105,504
85,828 98,656 106,590
95,443 111,051 120,780
81,511 96,077 107,063
80,161 89,707 92,374
76,743 82,470
64,854
103,929
110,649
110,107
124,499
109,505
107,855 111,901 112,388 111,727
113,833 116,691 116,055
113,159 113,243
124,352
Loss Reserving &amp;
Ordinary Least Squares Regression
A Love Story
Murphy variance assumptions
LSM
y  bx  e
y  bx  x e
y  bx  xe
Murphy: Unbiased Loss Development Factors
Or, More Generally
y  bx  x
 /2
e
The multivariate model
may be generally stated
as containing a
parameter to control the
variance of the error
term.
 is not a hyperparameter. Fitting using SSE will
always return  = 0
• Intuition
• Losses vary in relation to predictors
• Loss ratio variance looks different
• Observation
• Behavior of a population
• Diagnostics on individual sample
(Breusch-Pagan test)
In 2011, Glenn Meyers &amp; Peng Shi published
NAIC Schedule P results for 132 companies.
The object was to create a laboratory to
determine which loss reserving method was
most reliable.
12
10
88
Ln(MSE)
14
16
Log MSE - PP Auto
0
20
40
60
Company
80
10
55
00
Ln(UpperMSE)
15
Log UpperMSE - PP Auto
0
20
40
60
Company
80
-120000
-80000
-80000
Upper Error
-40000
-40000
00
Upper Error - PP Auto
0
20
40
60
Company
80
Observation of an Individual
Sample
Breusch-Pagan Test
• Use regression to diagnose
• Does the variance depend on
the predictor?
• Regress squared residuals
against the predictor
e   0  1 xi
2
i
An F-test
determines the
probability that
the coefficients
are non-zero.
12
10
88
Ln(MSE)
14
16
BP vs MSE - PP Auto
0.0
0.2
0.4
0.6
BP pVal
0.8
1.0
Caveats
Caveats
• Non-normal error terms render B-P
meaningless!
• Chain ladder utilizes stochastic predictors
• Earned premium has not been adjusted
Conclusions
• Breusch-Pagan test is not strongly
persuasive across the total data set.
• Homoskedastic error terms would support
unweighted calibration of model factors.
• Probably more important to test functional
form of error terms. Kolmogorov-Smirnov
etc. may test for normal residuals.
assumptions.
Test those assumptions.
Stop using models whose assumptions don’t
reflect reality!
Statisticians have been doing this for years. Easy
to steal leverage their work.