Predictive Modeling for Workers Compensation
36
Predictive Modeling for Workers Compensation David Otto, FCAS, MAAA EMB America LLC
Transcript of Predictive Modeling for Workers Compensation
Predictive Modeling for Workers
Compensation
David Otto, FCAS, MAAA
EMB America LLC
Modeling Workers Compensation Risks
OUTLINE
Plan Project
Gather Data
Build Models
Analyze
Implement
PURPOSE: To provide a technical discussion of solutions to the
challenges associated with modeling workers
compensation insurance.
Stage 5
Wrap-up &
Implement
Predictive Modeling Process
Stage 1
Plan
Project
Stage 3
Build
Models
Stage 4
Analyze
Impacts
Stage 2
Gather/Prep
Data
1. Plan Project: establish scope, objectives, and requirements
2. Gather & Prepare Data: gather data and create necessary model
files
3. Build Models: use historical data to build frequency and severity
models including underlying development models for medical and
indemnity (and possibly expense) losses and combine to form
modeled pure premiums
4. Analyze Impacts: analyze the renewal, competitive, and
profitability impacts of various proposals and finalize decisions
5. Wrap-Up: document decisions and communicate results
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Plan Project
Critical for proper project management
Understand objectives and goals of all stakeholders
– Timing
– End product (most often an underwriting score but how
is it expected to be used?)
• Tiering
• Schedule Rating – automated or recommended?
Ensure all stakeholders understand benefits of predictive
modeling over traditional techniques
Important to involve underwriting personnel in the process
because they will likely be the users of the final results
GOAL: Establish scope, objective, and
requirements
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Gather & Prepare Data
Data preparation can consume over half of the time spent on
a predictive modeling project
Key data challenges with a workers compensation project
– Volume
• Underlying distribution
• Dimensional dilution
• Exposure base
– Quality
• Policy/loss matching
• Null records
– Dimensionality
• Policy vs. claim rating variables
• External data
• Underwriting data
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Underlying Distribution
– Low frequency: fewer observations
Data Challenge: Volume
– High severity: greater volatility
Crunched Residuals (Group Size: 547)
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
0.000 0.002 0.004 0.006 0.008 0.010 0.012 0.014 0.016 0.018 0.020
Fitted Value
Two distinct clouds
imply Poisson
Diagonal spaces
imply Zero Inflated
Poisson
Capped severity
implies Gamma.
Propensity model
used to adjust for
censoring
Excess severity
models imply
Gamma or Inverse
Gaussian
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Data Challenge: Volume
Dimensional Dilution
– Observations are spread thinly over multiple states
– Extrapolating beyond niche markets
Indicated
Industry Rate
converts
categorical to
continuous
• Use indications from class codes with
higher volumes
• Use NCCI rate for class codes with low
volumes
• More years of data with a time element
adds credibility without sacrificing trend
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Class Codes: Ranked by NCCI Rate Policy Year
Data Challenge: Quality
Using new tools always seems to uncover previously undetected data problems that must be researched
Typical issues
– Bad data (e.g., 10 year old workers), especially for variables not used in rating
– Poor linkage between losses and policy and class characteristics
– No mapping of old groupings into new groupings (e.g., boundary changes)
– Inconsistency between variables
– Inconsistency within variables
– Free flow versus set level data capture
Issues above are magnified with long-tailed lines because of the need for a longer history of data
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Data Challenge: Quality
Null records prevalent in workers compensation data sets
– Creates complex aliases
Approved Program: (Approve/Not Approved) Loss Control Program: (Yes/No)
Data are the same records
in each dimension
Solution
– Create appropriate model structure to isolate the effect of the missing data
• Assign a parameter for unknown in one rating factor and group unknown with the base in all other rating factors
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Data Challenge: Dimensionality
Workers compensation systems capture small amounts of policy information and a large amounts of claim information
Policy DataFrequency
Claims DataSeverity
Data:
Class code
Territory
Experience mod
Minimum premium
Employee count
Credit score
Agency
Years renewed
Company size
Premium discount
Schedule credit
Injury type
Injury description
Injury location
Age
Gender
Attorney
Report lag
Weekly wage
Marital status
...many more
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Employee Manual
Subjective question:
Indicates presence
of manual NOT
effectiveness
Underwriting Acceptance Criteria
Underwriting Bias:
Detailed review of
risk alters the initial
rejection decision
Policy data is often supplemented with external data and underwriting information
– Incorporating external data
• Cost
• Time
• Maintenance
• Value
– Incorporating underwriting / policy application data
Data Challenge: Dimensionality
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Data Challenge: Dimensionality
Credit Score– Credit is a common external data source in personal lines
– Applicability to commercial lines?
– Problems:• Permanency:
– Personal lines: credit sticks with you for life
– Commercial lines: bad credit companies tend to go out of business and are often reborn under a new name
• Linkage– Personal lines: personal info (i.e. ss number) provides match
– Commercial lines: difficult to match
» Multiple mailing addresses
» Multiple company names within corporate structure
– Solutions:• Policy tenure may be a good proxy for credit
• Try linking data with multiple variables to increase success rate– Phone number
– Business name
– FEIN
Frequency
Severity
Build Models
Predictive modeling techniques are transportable to all types
of insurance but commercial lines modeling has its own
unique challenges that need to be properly addressed
Key modeling challenges with a workers compensation
project
– Components
– Development
Development
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Components
A common misperception in commercial lines is to perform
loss ratio modeling
Modeling loss ratios is significantly inferior to pure premiums
for both practical and theoretical reasons
– Habit from traditional pricing methodology or necessity
– Limitations of the predictive modeling software
– Short cut
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Components
On-level Premium
– Historical data must be re-rated using current rates (i.e.
extension of exposures)
• Each risk must be re-rated (average current rate level factors
will not work)
• Workers compensation requires REUNDERWRITING
– Credits/debits and special adjustments?
– Underwriting judgments will not be consistent
Extremely challenging task
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Components
Loss Ratio Distributions
– Tweedie
– Gamma
Heterogeneity within the response (frequency and severity
components) require dispersion modeling techniques
-3.5
-3.0
-2.5
-2.0
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
-1,000 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 9,000
Fitted Value
Stu
den
tized
Sta
nd
ard
ized
Devia
nce R
esid
uals
> .2
> .5
> .7
> 1.4
> 2.2
> 2.9
> 3.6
> 4.3
> 5.0
> 5.7
> 6.5
> 7.2
> 7.9
> 8.6
> 9.3
> 10.1
> 10.8
> 11.5
> 12.2
> 12.9
> 13.6
> 14.4
> 15.1
> 15.8
> 16.5
> 17.2
> 17.9
> 18.7
> 19.4
> 20.1
> 20.8
> 21.5
> 22.3
> 23.0
> 23.7
> 24.4
> 25.1
> 25.8
> 26.6
> 27.3
-5
-4
-3
-2
-1
0
1
2
3
-1,000 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 9,000 10,000 11,000
Fitted Value
Stu
den
tized
Sta
nd
ard
ized
Devia
nce R
esid
uals
> .3
> .5
> .8
> 1.6
> 2.5
> 3.3
> 4.1
> 4.9
> 5.7
> 6.6
> 7.4
> 8.2
> 9.0
> 9.8
> 10.7
> 11.5
> 12.3
> 13.1
> 13.9
> 14.8
> 15.6
> 16.4
> 17.2
> 18.1
> 18.9
> 19.7
> 20.5
> 21.3
> 22.2
> 23.0
> 23.8
> 24.6
> 25.4
> 26.3
> 27.1
> 27.9
> 28.7
> 29.5
> 30.4
> 31.2
Clusters of residuals due difference between frequency and severity signals
Dispersion modeling techniques improve fit
No real intuitive sense of what the scale parameter
response is in a dispersion model
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Components
Interpreting Loss Ratio Model Results
– No a priori expectation
– Shape of predictors will be dictated by how you are
priced across all types of risk
County Population Densities
Frequency Model
Can smooth out the
shape of the pattern
to highlight true
signal
Loss Ratio Model
Difficult to separate signal
from noise because
modeling premium
inadequacies
County Population Densities
Problem is magnified as you begin to introduce new variables
into the model
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Components
Implementing the results from Loss Ratio Model
– Black box
• Loss ratio modeling produces “black box” support
• Explanatory power of GLM’s are greatly diminished
– Buy-in from management, underwriting?
– Short cut
• Loss ratio model is not reusable
• Entire dataset must be re-rated each time analysis
is updated
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Development
Loss development within the predictive modeling process is not
necessary if one of the following conditions is met:
– Case incurred losses at the individual claim level are
reasonably accurate
• Common assumption in short-tailed lines
– Each claim at a given lag is assumed to develop by the same
proportion (i.e. the same development factor can be applied
to each claim)
• Common assumption in auto liability
– Basic limits losses
– Claims data matches policy data
For personal lines the time variable is sufficient
Workers compensation requires a more rigorous approach
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Development
Traditional loss development: aggregate all claims in each cell
within the historical triangle on a cumulative basis
1002005001100
1102005001075
02005001001
50002004000942
2001002004000834
4001202004000776
2003
2003
2003
2002
2002
2002
Accident
Year
2,000100 0 000443
400 400 125 000328
500 500 300 000124
250 0 0 0 000060
50 50 50 50 000021
5,000 1,000 1,000 0 000001
48 36 24 12 Claim
2005
2004
2003
2002
Accident
Year
210
1,320220
4,3251,425425
7,4502,1501,10050
48 36 24 12
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Development
Traditional loss development goals:
– Square the triangle
– Forecast the tail
Ultimate
10,324
11,729
14,987
7,450
48
2005
2004
2003
2002
Accident
Year
2,9791,162210
3,3851,320220
4,3251,425425
2,1501,10050
36 24 12
2005
2004
2003
2002
Accident
Year
210
1,320220
4,3251,425425
7,4502,1501,10050
48 36 24 12
Estimating aggregate reserves does not produce a solution for
allocating reserves to individual claims
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Development
GLM’s can be used to develop individual claims to ultimate
400362003000328
400242003000328
125122003000328
500362003000124
500242003000124
300122003000124
250482002000060
0362002000060
024 2002000060
0122002000060
50482002000021
50362002000021
2002
2002
2002
2002
2002
2002
Accident Year
5024 000021
5012000021
5,00048000001
1,00036000001
1,00024 000001
012000001
AmountAge Claim
1002005001100
1102005001075
02005001001
50002004000942
2001002004000834
4001202004000776
2003
2003
2003
2002
2002
2002
Accident
Year
2,000100 0 000443
400 400 125 000328
500 500 300 000124
250 0 0 0 000060
50 50 50 50 000021
5,000 1,000 1,000 0 000001
48 36 24 12 Claim
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Development
Development Challenges
– Functional relationship between variance and mean
• Poisson
• Binomial/Gamma process
• Zero Inflated model
– Functional relationship between characteristics and
incremental payment
• Multiplicative
• Additive
• Logit
• Non linear
– Squaring up the triangle on individual claims means that
reserves for IBNR claims are not estimated
• Workers compensation reporting lags have regulatory
guidelines
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Modeling Challenges: Development
Development Advantages
– Loss development enables the modeler to use claim
characteristics as a means to separate signal from noise
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Identify the pattern
with policy and
claim characteristics
WC Data:
Class code
Territory
Experience mod
Minimum premium
Employee count
Credit score
Agency
Years renewed
Company size
Premium discount
Schedule credit
Injury type
Injury description
Injury location
Age
Gender
Attorney
Report lag
Weekly wage
Marital status
...many more
Without claim characteristics, the underlying
signal will be difficult to identify
Modeling Challenges: Development
Once modeled, policy characteristics are then used to
translate the smoothed patterns using information available
when the policy is written
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Overall Mean“Best”Model
1 parameter for each
observation
Model Complexity
(Number of Parameters)
Sig
nal
No
ise
Analyze Impacts
Challenges:
Traditional qualitative tests tend not to be as conclusive in commercial lines:
– Standard error test leans towards underfitting the model
– Chi-Square tests leans towards overfitting the model• Tripped up on Unknown levels
Solution: Rely on additional diagnostics
– Balance
– Bias
– Lift
– Retrospective
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
RATING GROUP
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Class I Class II Class III Class IV
0
20
40
60
80
100
120
140
160
Rescaled Predicted Values - ccTota lSales
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
0%
2%
4%
6%
8%
10%
12%
14%
16%
18%
20%
22%
24%
0 5 10 15 20 25 30 35 40 45 50
Model
Prediction at
Base levels
Unknown is
significantly
different from
everything else
Analyze Impacts
Balance:
– Aggregate fitted results should be close to aggregate
observed data across data dimensions
Account Size: Out of balance, too low for smaller accounts
Aggregate
Observed
Aggregate
Fitted
Account Size: Used a more predictive model structure
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Aggregate
Observed
Aggregate
Fitted
Analyze Impacts
Bias:
– Compare observed versus fitted data across all accounts
-4
-3
-2
-1
0
1
2
3
4
5
6
11.2 11.6 12.0 12.4 12.8 13.2 13.6 14.0 14.4 14.8 15.2 15.6 16.0 16.4
Transformed Fitted Value
Stu
den
tized
Sta
nd
ard
ized
Devia
nce R
esid
uals
-3.5
-3.0
-2.5
-2.0
-1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
2.0
2.5
11.5 12.0 12.5 13.0 13.5 14.0 14.5 15.0 15.5 16.0 16.5 17.0 17.5 18.0 18.5
Transformed Fitted Value
Stu
den
tized
Sta
nd
ard
ized
Devia
nce R
esid
uals
Example: Overestimating
Underestimating
Overestimating
Example: No bias
– Similar to balance but now focus is on individual accounts
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Analyze Impacts
Lift
– Evaluate and quantify overall performance of selected
model structures
Model A
Model B
– Model B is producing a greater lift than Model A
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Analyze Impacts
Retrospective:
– Test performance on new data that was not used to build
model
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Actual inforce
loss ratio
Proposed Underwriting Score:
– Flat loss ratio from the latest inforce data set
– Volatility in loss ratio driven by immature data
Implement
Advanced techniques and technology enable the analyst to
look at more explanatory variables than previously imagined
– Results: Indications from predictive models will introduce
factors not currently used in rating and relativities for
existing factors that may be significantly different from
current relativities
– Dilemma: How to incorporate indications into existing
rating plan?
• Systems constraints
• Tradition
• Agency resistance
• Market perception
• Regulatory considerations
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Implement
Consistency between underwriting and pricing
– Underwriting emphasis is needed because of the uniqueness of
each exposure
– Overlap in qualitative assessments?
Moving from yes/no underwriting scores to rate refinement
– Underwriting scores help cherry-pick but do not establish the
right rate for every risk
– Either through rating or underwriting (example: tiers)
Buy-in of distribution channels
– Independent agents
– MGA’s
Sizing up the competition
– Industry focus on underwriting refinement makes it difficult to
evaluate competitors
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Implement
Scoring solutions allow multiple approaches toward implementation
– Accept/Reject criteria
– Tiering
– Schedule Rating
Tier #2 = 1.0
Tier #3 = 2.0
Tier #1 = 0.4
Underwriting Score: Captures difference between indications and manual rate
Schedule Rating
Debit/Credit
Acceptance
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Wrap-up
Create required filing support
Decisions made during the review will need to be
documented in accordance with actuarial
standards
Prioritize future actions
GOAL: Document decisions and communicate
results
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
Wrap-up
Success in the personal lines marketplace has been widely
publicized
Important to remember that transition to current level of
sophistication in the personal lines marketplace took many years
Can get a lot of value through small incremental steps
Implementation
Partial Rate
Adjustment
Data
ImprovementPredictive
Model
• Plan
• Implement
• Analyze
• Build Models
• Gather Data
America LLCMission:
EMB America seeks to help our clients solve complex problems and identify opportunities by providing the appropriate blend of value-added consulting and state-of-the-art software. In so doing, EMB America strives to develop long-term relationships with clients and be the consulting firm of choice for the business community.
For information:
Phone: 858.793.1425
Email: [email protected]
Website: www.embamerica.com
