average premium model - soa
TRANSCRIPT
Average Premium ModelActuarial Research Conference
Brant Wipperman, FCAS FCIA MSc (SFU)
July 27, 2010
Motivation
• Revenue Requirements
• Monitoring our book of business
1
2
Basic On-Level Average Premium
600
605
610
615
620
625
630
2002 2003 2004 2005 2006
Avg.
Premium
(
$)
Year
3
Personal TPL On-Level Avg. Premium
600
605
610
615
620
625
630
2005 2006 2007 2008 2009
Avg.
Premium
(
$)
Year
Overview
4
ExposureForecasts
AveragePremiumForecasts
PCA
Exposure Model
• Historical exposure data– Split into Personal and Commercial– Further split into vehicle use, location, and
bonus-malus groups
• An econometric regression model is fit to each group– Demographic– Economy
5
Vehicle Use Groups
• Personal
– Pleasure– Commute– Business– Senior– Motorcycle– Motor home– Collector
6
Location Groups
• Lower Mainland• Ridge Meadows• Fraser Valley• Squamish/Whistler• Pemberton/Hope• Okanagan• Kootenays• Cariboo• Prince George• Peace River• North Coast• South Island• Mid Island• North Island
7
Bonus-Malus Groups
• Claim Rated Scale
– Roadstar (43% discount)
– 25% to 40% discount
– 5% to 20% discount
– Base or surcharge
8
Overview
9
HistoricalExposure
Data
ExternalFactorsEconometric
Models FactorForecasts
ExposureForecasts
AveragePremiumForecasts
PCAExposure Model
Historical Exposure Data
• Too many groups for the average premium model
• Need a dimension reduction technique
• Want to keep all of the groups
• Linear dependencies exist
10
What is PCA?
• It transforms a number of correlated variables into a smaller number of uncorrelated variables
• Uses linear algebra
11
PCA Notation
12
)(1 ZZA Tn ⋅=
VVA λ=⋅
21−⋅= LVB
BZP ⋅= LBLVS ⋅=⋅= 2
1
TTTC ⋅=
Eigen Decomposition
• Linear algebra problem• Done on correlation matrix of
explanatory variables• Eigenvectors are new explanatory
variables (i.e. principal components)• Each associated eigenvalue represents
variability of eigenvector (or PC)
13
PCA Resolves the Issues
• Number of dimensions reduced• All groups ‘retained’• Linear dependencies eliminated
14
PCA Process
• Step 1: Create new set of explanatory variables
• Step 2: Determine how many new explanatory variables to retain
15
PCA
How many components?
0%
5%
10%
15%
20%
25%
30%
35%
1 2 3 4 5 6 7 8 9 1011121314151617181920212223242526272829303132
%
of
Variance
Principal Component
16
How many components?
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1 2 3 4 5 6 7 8 9 1011121314151617181920212223242526272829303132
%
of
Variance
Principal Component
17
88% 94%
Overview
18
Chosen PCs
HistoricalExposure
Data
ExternalFactorsEconometric
Models FactorForecasts
ExposureForecasts
Exposure ModelPCA
19
HistoricalExposure
Data30 correlated variables
Chosen PCs
Ortho-normaltransformation
PrincipalComponents 30 uncorrelated variables
ScreeProportion of varianceOther
6 uncorrelated variables
Overview
20
Chosen PCs
HistoricalExposure
Data
LinearRegression
Models
HistoricalAveragePremium
ExternalFactorsEconometric
Models FactorForecasts
ExposureForecasts
Chosen PCForecasts
AveragePremiumForecasts
PCAExposure Model
Average Premium Model
Modeled vs. Actual – Personal TPL
580
600
620
640
660
Jan-04 Jan-05 Jan-06 Jan-07 Jan-08 Jan-09 Jan-10 Jan-11
Avg.
Premium
(
$)
MonthModeled Actual
21
Modeled vs. Actual – Personal TPL
600
605
610
615
620
625
630
2004 2005 2006 2007 2008 2009 2010 2011
Avg.
Premium
YearActual 6 PCs 4 PCs 8 PCs
22
23
Recap - Advantages
• PCs uncorrelated
• PCs organized to reduce dimensionality
• Keeps most of original information
• Determine contribution of each variable
24
Recap - Disadvantages
• PCA process not familiar
• PCs can be hard to interpret
• PC weights may change upon updating
25
Is PCA Right For You?
• Does multi-collinearity roll off your tongue too easily?
• Are you confident in the set of explanatory variables?
• Do you want to reduce dimensionality without throwing away information?
• Have you been modeling for more than 4 consecutive hours?
26
For More Information
• CAS Discussion Paper
– PCA and Partial Least Squares: Two Dimension Reduction Techniques for Regression
• http://www.casact.org/pubs/dpp/dpp08/08dpp76.pdf