4. Predictive Modeling
▪Predictive Predictiveとは・・・
What is Modeling Modeling?
→ To sum up, it is “a technique to statistically project the future using
technique of data mining” .
What is “data mining”?
▪では「データマインニング」って何?
→ It is to get heuristic understanding of correlations and consequences
in the data that have not been recognized by exhaustively analyzing
large volume of data.
5. Predictive Modeling
“Data mining” has achieved the following success for example.
「データマインニング」は例えばこんな成果を挙げてきました。
An economist and a wine lover, Orley Ashenfelter, derived
the following theoretical formula, which calculates quality
of wine, by data interpretation.
Quality of wine = 12.145+0.00117×rainfall during winter+0.0614
×average temperature during growing period - 0.00386×rainfall during harvest period
→ Although it was ridiculed by wine critics who are “specialists in the field”
at first, it turned out to be able to project the quality of wine more
accurately.
Amazon’s “recommendation” function, Web advertisements
→ Data mining determines “a person who bought A and B should be
interested in C”, where the data of products bought and pages clicked in
the past of a page viewer has been accumulated.
6. Predictive Modeling
To identify relationship of the data that had not been recognized by
analyzing large volume of data and utilize it for a certain purpose for
the field of insurance, too.
Many companies, particularly in US, have introduced data mining tools
for P&C area.
Application of data mining for various life insurance purposes is
expected, too.
– Policies with what kinds of characteristics tend to be surrendered
– Potential customers with what kinds of characteristics tend to purchase
what kinds of insurance
– To identify risk of selling limited declaration type and/or no screening type
– To differentiate products
– To achieve lower premium for preferred risk
– For cost down and speed up of screening
– To accurately project trend of mortality and/or morbidity rates
– ・・・・
7. Predictive Modeling
Example of application
In-force business
Product profit
Marketing Underwriting management
development
section (UW) section and risk
section management
section
8. Predictive Modeling
Example of application for marketing section
Product In-force business
Marketing Underwriting profit management
development
section (UW) section and risk management
section section
Traditionally,
– Acquisition of policies by expanding potential customer base through connection of
sales reps and/or consulting sales and other communications by the staff and
planners
– Acquisition of policies with less strict screening through DM and/or website,
marketing to particular corporations, bank-counter sales
– ….
Predictive Modeling allows cost reduction through automatic selection of
distribution channels of new products and/or automatic segmentation of DM
recipients, and competitive superiority by offering business for segments that
had not existed. It can also be utilized for effective training of sales
representatives.
9. Predictive Modeling
Example of application for UW section
In-force business
Product profit management
Marketing Underwriting
development and risk management
section (UW) section
section section
Ordinary UW includes declaration of current health condition, history of sickness and
occupation, interview of a reviewer or a doctor, submission of health examination report,
and sometimes blood tests in case of high-end products, that requires cost. As it also
needs some time especially for high-end products, potential customer may loose his/her
appetite to buy or move to other firms during the waiting period.
Comparison of mortality rates
Predictive Modeling can Predictive Model vs. Full Underwriting Image of UW cost reduction
automatically identify a
group of people with high
risks. It is expected to be
used to simplify UW, or
specify a group that needs
whole process of UW.
※Charts are referenced from Reference material 2.
10. Predictive Modeling
Example of application for product development section
Product In-force business
Marketing Underwriting profit management
development
section (UW) section and risk management
section section
Traditionally, products have been developed by offering lower premium
for younger age groups and/or segmenting premium for smokers.
Predictive Modeling is expected to be utilized for product development
which offers competitively priced products and those for cross-sales
with segments that are different from traditional classifications such as
health condition and/or smoking status.
11. Predictive Modeling
Example of application for in-force business profit management and risk management
section
Product In-force business
Marketing Underwriting profit management
development
section (UW) section and risk management
section section
In-force business risks shall change (decline of selection effect, anti-
selection, move of standard risk to preferred risk, etc.)
Predictive Modeling is expected to be utilized to understand what risks
are held in certain groups of in-force business, what groups tend to be
surrendered, and to what groups a retention program should be
focused.
12. Predictive Modeling
Representative methodologies of Predictive Modeling
GLM
Statistical methods
Bayesian statistics
Generic Algorithm
Non-statistical methods Neural Network
(Machine Learning) Decision Tree
Scoring
Recent study reveals that use of a customized methodology suitable for a
purpose shows better results than those using a certain methodology only.
13. Predictive Modeling
Image of Decision Tree
To mechanically develop a tree
Using a tree
suitable for a purpose
Single
Yes No
Income of over
Segment 2 X million yen
Yes No
Segment 3 Purchase at banks
Yes No
Segment 4 ····
14. Predictive Modeling
Challenges for introduction Company data External data
– Model development techniques
– Administration structure
• IT
• Reform of corporate culture
• Training of internal staff having literacy Predictive Modeling
– Data to be used Algorithm
To start using supplemental purposes
rather than replacimg the existing
methodology at once
To receive a tool and support from a
vender On-line UW system, sales reps., accounting department,
To use free tools (such as Weka and R) etc.
To gradually accumulate the company
data, while using the external data as a
source data at first, as the company Existing methodology
data is ideal.
15. PREDICTIVE MODELING
REFERENCE MATERIALS
1. “Super Crunchers” by Ian Ayres, translation by Hiroo Yamagata
published by Bungeisyunjun
2. “Predictive Modeling for Life Insurance”, Deloitte Consulting LLP,
April 2010
3. “Predictive Modelling for Commercial Insurance”, General Insurance
Pricing Seminar, 13 June 2008 London
4. Free design material site “来夢来人”
http://www.civillink.net/fsozai/illust4.html
17. Complexity Science
A technique applying a way of consideration of Complexity Science
(complex type), which has long been discussed in other areas of
science, such as social science, has started to be discussed recently
in the insurance industry.
A point that it is different from the traditional concept at first is
concept of “Agent-Based Model” below.
Agent-Based Model:
“A concept, which explains complex phenomena by behavior of a
system came into effect by a relationship of Agents having a certain
characteristic and a simple behavior pattern. It is a ‘concept’ rather
than a certain model or a formula.”
18. Complexity Science
▪ Understanding of it by the conventional system:
Input Complex relational expression Output
with multiple variables
▪ Understanding of it by the system of Agent-Based Model:
Environment
Relationship
Agent
4 Output 1
Agent
Agent 3
Input 1 1
Agent Output 2
Input 2 Agent 5
2
19. Complexity Science
Example
Agent ⇒ Airport
Relationship ⇒ Airline route
Network of an airline route can be
expressed as a graph.
“Network Science” described later
discusses how airport and airline route
should be increased to develop a network,
which is least affected by an accident,
where an airport becomes unavailable,
based on the features of the graph.
※ Charts are referenced from Reference material 2
20. Complexity Science
▪ Why is Agent-Based Model necessary?
なぜAgent-Based Modelが必要か?
– It allows “explaining” an actual complex system in more realistic way.
→ Once the system is understood,
• It is easy to explain to the concerned people.
• It allows real-time crisis management.
• Important data needs to be focused on is identifiable.
21. Complexity Science
Agent-Based Model · · ·
+
Concept of “network” Example of Cellular Automata: Forest fire
→ “Network Science”
Concept of
+
“free behavior of an Agent”
→ “Cellular Automata”
+
Concept of development of
“environment”
→ “Artificial Societies”
+
“Involvement of a model user”
→ “Serious Game” Example of Artificial Societies: Sugarscrape
※Charts are referenced from Reference material 2
22. Complexity Science
How to utilize it for insurance
Case 1: “Surrender behavior simulation” (Cellular Automata)
(2002, Charles Boucek and Thomas Conway)
Agent: Policyholders, insurance company (itself, competitor), agents, etc.
Relationship: Policyholder and insurer, agent contract, etc.
Assumptions:
– Policyholder shall age per elapse of time.
– Health condition and marital status shall also change.
– Policyholder shall surrender a policy or replace a policy to other company’s per
some criteria.
→ It allows simulations of how the distributions of policyholders would
change after certain period of time assuming if the company keeps the
level of premium and if it raises its premium, and how they affect the
profitability.
23. Complexity Science
How to utilize it for insurance (continued)
Case 2: “Epidemic simulation” (Artificial Societies)
(2003 – 2004, Epstein)
Agent: People who live in town
Relationship: Infection of epidemic
Environment: Home, workplace, school
Assumptions:
– A person goes to workplace or school.
– A person would be infected at random, if
he/she stays at the same place, where an
infected person stays.
→ It turns out that approaches such as “vaccinating all the hospital
workers”, “allowing second vaccination as an option” and “isolating
infected patients in hospitals” are more effective than programs actually
taken in the past by reiterating simulation of situations introducing various
prevention programs.
※ Charts are referenced from Reference material 2
24. Complexity Science
How to utilize it for insurance (continued)
Case 3: “ERM: Risk mapping for risk appetite”
(2011, N. Allan and others)
Agent: Risk factor (probability
variable)
Relationship: Cause and effect
(occurrence probability)
Assumptions:
– To map correlations of risk
factors using a concept of
Bayesian Network and
calculate occurrence
probability allocated to each
factor.
→ Complicated risks such as operational risk can be evaluated upon
recognized by a “natural” system.
※ Charts are referenced from Reference materials 1 & 3.
25. Complexity Science
Simple question
▪素朴な疑問:
– “Isn’t it just a simplified game without any reality after seeing some cases?”
→ Although there is such an aspect, it helps us to “understand” the types of impact
on the system by changing some parameters.
▪結局のところ、私の考える複雑系のコアの利点とは core is as follows:
In short, I think advantage of complex system
– It allows “explaining” an actual complex system in more realistic way.
(reprint)
→ Once the system is understood,
– It is easy to explain to the concerned people.
– It allows real-time crisis management.
– Important data needs to be focused is identifiable.
26. Complexity Science
Challenges for introduction
▪導入にあたっての課題
– Hardest point is “what does complex system methodology help is not
very clear”.
– How to apply a model to an actual issue?
– There is no one tool for all purposes.
• There are variety of tools to support development of Agent-based Model
(ABM).
Please see “Tools for Agent-Based Modelling”.
(http://www.swarm.org/wiki/Tools_for_Agent-Based_Modelling)
Please join us at discussion forum and other
opportunities, if you are interested in. We are still
creating real cases of applications by training
interested individuals as specialists.
27. Complexity Science
Reference materials
1. “Applications of complexity science to risk appetite and emerging
risk” Neil Cantle, Milliman, Yorkshire Actuarial Society
2. “Complexity Science – An introduction (and invitation) for actuaries”
Prepared by Alan Millis, June 10, 2010
3. “A review of the use of complex systems applied to risk appetite and
emerging risks in ERM practice” N. Allan, N. Cantle, P. Godfrey and Y.
Yin, Presented to The Institute and Faculty of Actuaries, 28
November 2011 (London)