... for thought – carefully constructed manual factor combinations can also yield
material benefits .... LOOK. 0. 5. LOOK. 5. FIAT PUNTO 75 SX. FIAT PUNTO 75.
What segments can I dream up? Y Young mothers th with ith 4x4s 4 4 doing d i the school run Old people l driving di i a Honda H d Jazz J going slowly in front of me
31
Key y messages g
– “Saddle” method of interaction detection can identify many subtle interactions and yyield materiallyy more p predictive models – This is no replacement for thought – carefully constructed manual factor combinations can also yield material benefits
32
Agenda g
Enhanced GLMs
and vehicle grouping
Duncan Anderson
Sami Abdel-Gadir
33
The starting g point p – the ABI 50 vehicle classification
•
New vehicles classified according to: – Damage and parts costs – Repair times – New car values – Performance – Security
•
50 groups in use plus suffixes
•
Imported cars and specialised purpose vehicles e e.g. g kit cars are not classified
•
For details see: http://www.thatcham.org/abigrouprating/index.jsp?page=429 34
How g good is ABI 50 for risk models and pricing? p g •
Useful benchmark
Proportion of TPBI and TPPD Burning Cost - market measure 100% 90%
•
Public awareness
80% 70%
•
Very good predictor of total loss?
60% 50%
• •
Good predictor of claim frequency? Better predictor of AD claims experience than TP?
is a one-size fits all vehicle group th b the bestt option? ti ?
1
•
35
Insurer classifications
36
Postcoding g - framework
Standard Policy F t Factors
Random Noise
External Geographical Factors
Residual R id l Spatial Variation
37
Spatial p smoothing g
– Credibility family method – Can adopt distance based or adjacency based approach
Unsmoothed
Residual Spatial Variation
Distance
Adjacency
Smoothed 38
Car classification – translating g the framework
Standard Policy F t Factors
Random Noise
External Vehicle Factors
Residual R id l Spatial Variation
39
Back to basics
Dimensions
Body style
Safety
Performance
Security
Cost
Brand Appeal
Use 40
Body y style y classification It’s hard! – No universally adopted system in place – Many variants to classify – New bodystyles have emerged – Some vehicles attempt to defy classification Hatchback
Cabriolet
Spider
Saloon 41
Car classification – translating g the framework
External Vehicle Factors Performance Di Dimensions i Safety Security Costs Bodystyle Classification
Standard Policy Factors
Random Noise
Residual Spatial Variation
Vehicle Weight Exposure
TPPD Frequency
42
Car classification – translating g the framework
Residual Spatial Variation Standard Policy Factors
Random Noise
Unsmoothed
Smoothed
Requires q a vehicle space... p
External Vehicle Factors
43
Example p adjacencies j
0
PEUGEOT 206 PEUGEOT 206 ZEST STY LE ZEST STYLE
VOLKSWAGEN VOLKSWAGEN POLO E SDI POLO E SDI
5
0
5
PEUGEOT 206 PEUGEOT 206 LOOK LOOK
0
5
FIAT PUNTO 75 FIAT PUNTO 75 SX SX
0
TOYOTA YARIS TOYOTA YARIS L ZINC D-4D L ZINC D-4D 5
44
Putting g it all together g Classifying existing cars
– Techniques learned from postcode classification can be successfullyy applied pp to car classification – There are practical challenges around making best use of external data but these can be overcome