The Impact of
Predictive Analytics
on Reserving
Brian Z. Brown, FCAS, MAAA
Principal and Consulting Actuary
Stan Smith
Predictive Analytics Consultant
April 29, 2016
Advancements In Reserving
Use of stochastic methods. Advancements in computing power have allowed for
more sophisticated reserving methodologies
 Distributions of loss reserves to better understand liabilities
Deterministic methods still dominant. Traditional development methods are still the
go-to choice for most reserving actuaries
 Methods are easier to explain to decision makers
Leaving something on the table? With the abundance of data being captured, could
predictive analytics help us better understand our claims data in order to make better
reserve projections?
2
Advantages of applying analytics in reserving
Strategic Reasons.
 Competitive advantage
Operational Reasons.
 Claims management/operations
 Better information for pricing
 Better information for reinsurance procurement
Financial Reasons.
 Quicker closure rates (lower claims costs)
 Lower defense costs
 More accurate case reserves on higher value claims
 Early intervention typically reduces claim costs
3
Predictive models
for case reserves
4
5
Insurers cannot rest
 “Companies that have not actively invested in improving
their pricing sophistication, efficiency and risk management
are at a competitive disadvantage and will not be relevant
in the long term”.
 Source: A.M. Best
6
MONEYBALL AND SABERMETRICS
Indicators of Offensive Success
MONEYBALL
MEASURES
TRADITIONAL
MEASURE
Home Runs
Batting Average
Stolen Bases
RBI’s
On-Base %-age
Slugging %-age
Pitch Data
Expected Future Runs Scored in an inning given
certain conditions. (1961-77 data set)
7
Winning an unfair game
“People operate with beliefs and biases.
To the extent you can eliminate both and replace them with data,
you gain a clear advantage.”
Michael Lewis,
Moneyball: The Art of Winning an Unfair Game
8
Analytics can help identify “Useful” data
 Leverage more of the data being captured
Traditional Approach
Analyzed
information
Big Data Approach
All available
information
analyzed
All available
information
Analyze small
subsets of data
Analyze all data
9
Supervised versus Unsupervised approaches
10
Text mining variables
 Text mining refers to the process of deriving
relevant and usable text that can be parsed
and codified into a word or numerical value.
 Text mining can identify co-morbid conditions
and/situations that will have profound impact
on the outcome of a claim.
CXR
smoking
Pain
unchanged
SAMPLE KEY WORDS/PHRASES









Diabetes/insulin/injections
Packs day/coughing
Pain killers/anti-depression
Children/school
Pain unchanged
Height/Weight
Homemaker wife went to work
c/o, CXR, FB, FX
CBT – Cognitive Behavior Therapy
Text sources: Adjuster notes, medical
reports, independent medical exams, etc.
11
Modeling architecture
 Data Store. All historical data collected and organized.
 Training. Identifying company/internal/external data specific patterns.
 Testing. Using “hold out” sets to measure the accuracy of predictions.
12
Virtual data warehouse
OPERATIONAL SYSTEMS
• Improved
Decisions
UNDERWRITNG
STRATEGIC
• More Timely
Decisions
NEW Business
• Efficient Use of
Resources
ETL
LOSS DATA
Medical Vendor
Data
TACTICAL
Other
DATA
WAREHOUSE
SCORING
ENGINE
Milliman DB
3RD PARTY
• Improved
Results
BUSINESS
INTELLIGENCE
• REPORT
• MONITOR
• MEASURE
Closed Loop
Reporting
DATA ENRICHMENT
13
Complementary Analytics Solutions
Underwriting
Better Decisions
WC Claim
Better Outcomes
• Leveraging data and analytics
• Improved pricing and
segmentation
• Improved client targeting
• Predictive Modeling - proactive claim management
• Data Warehouse comprehensive view of internal
data
14
Decision support example - claims
Quickly identify “creeping catastrophic” claims.
 Less than 20% of claims cause 80% of losses
Create better claims outcomes with more timely and more detailed
information.
 Loss cost reductions that generally range from 3-6% per year
“Operationalize” into claims/medical protocols/rules.
Integrate management of all available sources of data/information.
“Second pair of eyes” on existing claim/medical vendors.
Ancillary benefits.
 Data driven culture
15
Segmentation analysis
Divide All Claims into 5 buckets of 20% each.
 After Scoring distribute by Risk Score
 Highest Risk to the Right
 Lowest Risk to the Left
 Each Claim has a individual score
 Worst Claim far right vs. Best Claim far left
 Then add actual losses to test model accuracy
20%
20%
20%
20%
20%
16
Predictive modeling in action
Early ID < Day 30
High Risk
 Models Identify 20% of Claims
that have 78% of total costs
Medium Risk
Low Risk
11.47%
2.19%
3.15%
4.74%
17
18
Actuarial Dashboard – Auto Liability
19
Actuarial Dashboard – Incurred Loss & ALAE
20
Actuarial considerations
Changing development patterns. Nothing new for reserving actuaries, but important
to monitor and be aware of implications.
 Reserving
 Pricing
 Reinsurance
 Communication with different business units is important
21
Initial measurements
Is the model identifying problematic claims earlier?
Impact of early intervention. Cost savings.
Impact on average case reserves.
 Reserve strengthening
 Adequate
Impact on payment pattern. Early intervention may result in high payments initially but
lower total payments for a claim.
22
Actuarial considerations – after implementations
Adjustments to standard reserving methods
 Berquist-Sherman
Varying / Adjusting assumptions
 Loss Development Factors based on
 Anticipated patterns to change
 Measurement of how patterns are changing
 Expected Loss Ratios based on
 Projected cost savings
23
Predictive model
case studies
Closed Claim Impact
Client Case Study #1
25
Payment Trends – Totals over 10 Year Period
26
Payment Trends – Averages over 10 Year Period
27
Claim Age & Payment Trends – Model Impact
28
Reserving Impact
Client Case Study #2
29
Reserving Trends – Model impact first 30 days
30
Reserving Trends all open claims – Model Impact
31
Actuarial considerations
Some models establish an ultimate per claim
 Difference between model ultimate and incurred losses produces an indication
of IBNR
 How is pure IBNR determined?
 Is the model ultimate an additional methods used for reserving?
 How does the model ultimate change over time?
Documentation. Necessary so that another actuary may be able to review work product.
 ASOP’s
32
Implementation
considerations
33
Implementation planning is critical
Who has access to the model and its results? Access to the model and its output
could potentially influence behavior
 Could impact how case reserves are set in the future if claims adjusters have
access to the model output
 Impacts development patterns used for traditional actuarial reserving and ratemaking
techniques
 Would model still be relevant or applicable if underlying inputs are being changed?
Staff morale. Be cognoscente of how various business units might react to the
implementation of the model
 Intentions of model should be communicated to those who might be affected by
the model
 Fear of job elimination (claims adjusters, traditional reserving actuaries)
How to measure return on investment? Predictive modeling can be costly; make
sure you’re getting the most out of your investment
 Designate benchmarks before modeling
Regulation. Will regulators at some point need to review model?
34
Thank you
Brian.Brown@Milliman.com
Stan.Smith@Milliman.com