A Public Policy Prac tice Note

Credibility Practice Note

July 2008

American Academy of Actuaries’

Life Valuation Subcommittee
 Credibility Practice Note

The American Academy of Actuaries is a national organization formed in 1965 to bring together, in a
single entity, actuaries of all specializations within the United States. A major purpose of the
Academy is to act as a public information organization for the profession. Academy committees, task
forces and work groups regularly prepare testimony and provide information to Congress and senior
federal policy-makers, comment on proposed federal and state regulations, and work closely with the
National Association of Insurance Commissioners and state officials on issues related to insurance,
pensions and other forms of risk financing. The Academy establishes qualification standards for the
actuarial profession in the United States and supports two independent boards. The Actuarial
Standards Board promulgates standards of practice for the profession, and the Actuarial Board for
Counseling and Discipline helps to ensure high standards of professional conduct are met. The
Academy also supports the Joint Committee for the Code of Professional Conduct, which develops
standards of conduct for the U.S. actuarial profession.
This practice note is not a promulgation of the Actuarial Standards Board, is not an actuarial
standard of practice, is not binding upon any actuary and is not a definitive statement as to what
constitutes generally accepted practice in the area under discussion. Events occurring
subsequent to this publication of the practice note may make the practices described in this
practice note irrelevant or obsolete.

This practice note was prepared by the Life Valuation Subcommittee of the Academy of Actuaries.
The Academy welcomes your comments and suggestions for additional questions to be addressed by
this practice note. Please address all communications to Natalie Jones, Life Policy Analyst at
[email protected].

The members of the work group that are responsible for this practice note are as follows:
Robert DiRico, ASA, MAAA, Chair
Donald Behan, FSA, MAAA, FCA, Ph.D.
Robert Buzecan, FSA, MAAA
Mike Davlin, MAAA
Steven Ekblad, FSA, MAAA
Yu Luo, ASA, MAAA
Susan Miner, FSA, MAAA
Tomasz Serbinowski, FSA, MAAA, Ph.D.
Randall Stevenson, ASA, MAAA
Joth Tupper, FSA, MAAA
Van Villaruz, FSA, MAAA
Ali Zaker-Shahrak, FSA, MAAA, Ph.D.
Trevor Zeimet, FSA, MAAA

The work group would also like to thank the following for the valuable contributions:
Stuart Klugman, FSA, Ph.D.
Thomas Herzog, ASA, Ph.D.
 TABLE OF CONTENTS

1. Introduction

1.1. The Purpose of a Practice Note

1.2. Scope
1.3. Current References to Credibility Theory

2. Regulatory Guidance

2.1. Examples of Credibility Standards Used by State Regulators

2.2. Credibility Theory and Regulatory Practice – An Example: Credibility of Annual
Claims Experience Data as Reported by Health Insurance Companies to State
Regulatory Agencies

3. Industry Practice

3.1. A Limited Fluctuation Credibility Example

3.2. A Greatest Accuracy Credibility Example
3.3. Credibility for Mortality Ratios
3.4. Practical Considerations for Credibility in Reinsurance Pricing
3.5. Practical Considerations for using Credibility for Mortality
3.6. An Example Using Limited Fluctuation Credibility to Update Lapse Assumptions
3.7. Credibility Conflict I
3.8. Credibility Conflict II

4. A Short History and Background of Credibility Theory and its Usage in Actuarial Practice

5. Appendices

5.1. Bibliography of Recommended Reading on the Subject of Credibility Theory and

Related Approaches
5.2. Regulatory Guidance

Credibility Practice Note, July 2008

Page 2 of 55
1. Introduction

This practice note was prepared by the Credibility Practice Note Work Group of the Life
Valuation Subcommittee of the American Academy of Actuaries. It is not a promulgation of the
Actuarial Standards Board, is not an actuarial standard of practice, is not binding upon any
actuary and is not a definitive statement as to what constitutes generally accepted practice in the
area under discussion. Events occurring subsequent to this publication of the practice note may
make the practices described in this practice note irrelevant or obsolete.

1.1. The Purpose of a Practice Note:

The purpose of practice notes is to provide information to actuaries on current or

emerging practices in which their peers are engaged. They are intended to supplement
the available actuarial literature, especially where the practices addressed are subject to
evolving technology, recently adopted external requirements, or advances in actuarial
science or other applicable disciplines (e.g., economics, statistics, or enterprise risk
management). Practice notes are not interpretations of actuarial standards of practice
nor are they meant to be a codification of generally accepted actuarial practice.
Actuaries are not in any way bound to comply with practice notes or to conform their
work to the practices described in practice notes. (Guidelines for the Development of
Practice Notes, as adopted by the American Academy of Actuaries Board of Directors
September 25, 2006)

1.2. Scope

This practice note is intended to provide information on common practices and

approaches related to actuarial issues for which credibility theory or a related approach
may be applied, such as:

- Determining assumptions to use in modeling company cash flows. This may involve
the blending of company data with standard/industry tables or data.
- Determining the level of reliance that can be placed on company experience (e.g., in
the rate making or reserving process).

There are certain things that should be understood about this practice note in order to
properly apply credibility theory or a related approach to solve a business issue:

- The practice note provides illustrative examples of how credibility theory and related
practices have been or may be applied. Proper analysis requires the individual to
understand the similarities and differences between these examples and their specific
business issue. There should be a sound rationale for application of any of these
approaches and an understanding of why other approaches may not be appropriate.
- The examples presented in this document do not represent the official position of any
company, regulatory body, or the American Academy of Actuaries (Academy).
- The examples presented illustrate some methodologies that can used to solve the
problem and potential strengths or weaknesses of that approach.

Credibility Practice Note, July 2008

Page 3 of 55
- This practice note does not advocate a ‘one size fits all’ approach to solve a given
business issue.
- Section 4 provides useful background information that the reader, with a beginner’s
knowledge of credibility, may find helpful in understanding the examples that are
presented in the main body of the practice note.
- Finally, the Appendices, in Section 5, provide a bibliography of suggested materials
in order to gain a more complete understanding of this topic. The reader is strongly
encouraged to use these resources for their personal and professional benefit.

Credibility Practice Note, July 2008

Page 4 of 55
1.3. Current References to Credibility Theory

The use of credibility theory within publications that provide guidance to actuaries has
increased recently with the development of principle-based approaches for life and
annuity reserve and capital calculations. Prior to this, credibility theory was more
commonly referenced in property and casualty documents.

Examples of how credibility is referenced in current actuarial literature are shown below:

ASOP 25, Credibility Procedures Applicable to Accident and Health, Group Term Life,
and Property/Casualty Coverages

“The purpose of this standard of practice is to provide guidance to actuaries in

the selection of a credibility procedure and the assignment of credibility values to sets of
data including subject experience and related experience. . . This standard of practice is
applicable to accident and health; group term life; property/casualty coverage; and other
forms of non-life coverage.”

American Academy of Actuaries Report on Setting Regulatory Risk-Based Capital

Requirements for Variable Annuities and Similar Products (C3P2), June 2005

This report was subsequently adopted by the NAIC with little change. The use of
credibility theory is evident throughout the document:

• Prudent Estimate assumptions are to be made based in part on the degree of

credibility of the company’s underlying data.
• The approach used to adjust the mortality curves for the credibility of the
underlying data should be “suitable for credibility.” The document also states that
“when credibility is zero, an appropriate approach should result in a mortality
assumption consistent with 100% of the statutory valuation mortality table used in
the blending.”
• The document further states that “The credibility procedure used shall . . . contain
criteria for full credibility and partial credibility that have a sound statistical basis
and be appropriately applied..” and “Documentation of the credibility procedure
used shall include a description of the procedure, the statistical basis for the
specific elements of the credibility procedure, and any material changes from
prior credibility procedures.”

The January 2008 NAIC draft of VM-20 (Valuation Manual, Section 20), providing
principle-based valuation guidance for life products, states that credibility theory is to be
used to determine the appropriate mortality assumption and that the prudent estimate
assumption for each risk factor shall be based on available, relevant and credible
experience.

Credibility Practice Note, July 2008

Page 5 of 55
Further references to credibility theory can be found in the Group Term Life Waiver of
Premium Disabled Life Reserves Actuarial Guideline, the Medicare Modernization Act
and the Medicare Part D Practice Note from the Academy.

Credibility Practice Note, July 2008

Page 6 of 55
2. Regulatory Guidance

2.1. Examples of Credibility Standards Used by State Regulators

There is no single credibility standard or approach used by all states. Appendix 5.2 provides
excerpts from six state insurance regulation manuals describing how to apply credibility
theory or a related approach to solve insurance problems. A summary of Appendix 5.2 is
provided below.

• 5.2.1 – This section of regulatory guidance describes how Florida expects HMO
pricing assumptions to be developed using credibility theory. It provides full
credibility to plans that have 2,000 or more subscribers in force. It does not offer
any reasoning for the selection of the 2,000 value over other possible values.

• 5.2.2 – This section of Colorado guidance is very similar to that above. It uses
2,000 lives and 2,000 claims per year for full credibility. It applies to all health
insurance rate filings. It further provides a partial credibility formula based on the
square root of the actual experience divided by the full credibility standard.

• 5.2.3 – This section of North Carolina guidance uses 1,082 as the value for full
credibility. This sample provided guidance for credit life accident and health
policy rates.

• 5.2.4 – This section of Texas guidance provides credibility standards for Medicare
Supplement policies. This standard ascribes full credibility with 2,000 or more
lives, no credibility at 500 or fewer lives and linearly interpolates between 500
and 2,000 lives to give partial credibility.

• 5.2.5 – This section of California guidance applies to group life and disability
rates. This section provides a credibility table that sets full credibility for group
life plans at 40,000 lives and for disability plans at either 3,125 or 4,651
depending on the type of plan.

• 5.2.6 – This section of Maine guidance provides a table of values used to calculate
the credibility factor used to adjust premium rates. The credibility factors are
based on number of life years insured and claim count.

A common element of these regulations is the setting of a credibility threshold in terms of the
number of lives insured. These samples are appealing because of their simplicity and ease of
use, but they do not offer justification for the values chosen nor the basis upon which their
particular formula was chosen over other possible approaches.

Credibility Practice Note, July 2008

Page 7 of 55
2.2. Credibility Theory and Regulatory Practice - An Example: Credibility of Annual
Claims Experience Data Reported by Health Insurance Companies to Regulatory
Agencies

BACKGROUND

Health insurance policies sold to individuals are subject to state regulation. Companies
selling such policies need to get prior authorization for the premium rates they charge and for
future rate increases. States regulators generally require insurance companies selling
individual health insurance policies to set their premium rates at a level that is considered to
be “reasonable.” Reasonableness of premium rates is judged by the loss ratio that they
produce.

The standard of reasonableness is often expressed in terms of loss ratio being above certain
minimum levels. In California, for example, the standard of reasonableness for major
medical expense policies is expressed in terms of future and lifetime loss ratios being above
70 percent.

The loss ratio for any given period is the ratio of incurred claims to earned premiums in that
period. Since it is not difficult to forecast Per Member Per Month (PMPM) premium
income, the variability of the loss ratio for a given period will be due to the variability of the
claim costs in that period.

Companies attempt to forecast future claim costs using two sources: external industry data
and the company’s own historical PMPM incurred claims costs data. They generally combine
the information obtained from these two sources to arrive at their “best estimate” for the
future period. For example, if external sources indicate that PMPM cost next period will be
20.0% higher than its current level, and the company’s own experience indicates that PMPM
claim costs will increase by 30.0%, the company might report a 27.0% anticipated claim cost
increase and file for a similar premium rate increase.

The key issues that the company actuaries and regulators face are the following: To what
extent is it right and proper to forecast future claim cost by extrapolating a company’s
historical data? If we conclude that the experience data is not a completely reliable source
for this exercise, what other sources should be used, and in what way?

CREDIBILITY THEORY

Forecasting future PMPM claim costs is a statistical estimation problem with a twist. In a
standard estimation problem the actuary is given the parameters of the distribution and asked
to forecast the outcome of some future trial. For example, if the distribution happens to be
normal, the value of mean, μ, is given and standard deviation, σ, and the issue is to forecast
the outcome of a drawing, or average value of a number of drawings, from this distribution.

Credibility Practice Note, July 2008

Page 8 of 55
Even when ignorant of the value of the parameters of the distribution, the actuary usually has
access to some “best estimates” of these parameters, and can use them in the estimation or
forecasting exercise.

In attempting to forecast future PMPM claim costs, the actuary does not know the value of
the parameters of distribution. Worse still, there are many possible, and distinct, values that
present themselves as “the best estimates” of the underlying parameters. One of these values
is the one that occurs if the company’s recent historical experience is viewed. The others are
the values that are reported in various industry studies of claim costs of policyholders who
are considered to be “similar” to the policyholders of the company for which the actuary is
trying to forecast the PMPM claim costs. How can the actuary select a “best estimate” to be
used in the forecasting exercise and how can they justify their selection procedure?

“Credibility Theory,” as discussed in actuarial textbooks, attempts to resolve the problem by

a compromise solution: Rather than choosing one or the other “best estimate,” why not
choose a value that is a linear combination, or weighted average, of the best estimates?
Credibility theory is employed as follows: As a preliminary step, choose from the
multiplicity of outside or industry estimates the one that is judged to be the closest in
characteristics to the block of business for which the actuary is attempting to forecast future
PMPM claim costs. Next they combine the estimate obtained from the company’s experience
with the “best” outside estimate, using some appropriate weights, to arrive at a forecast of
future claim costs. The weights are determined as a function of the number of policyholders
that the company has had in the most recent historical period. The basic idea is to give more
weight to the company’s historical experience as the block of business in question is larger.

In detail, using credibility theory, the actuary can forecast future PMPM values as follows:
start with a PMPM claim costs estimate derived from company’s most recent historical
experience - μcomp – and an estimate of PMPM claim costs derived from “the most
appropriate” external source - μind. Next determine two values, N0, and NF. N0 is a minimum
sample size; if the company’s claim costs are derived from a sample size that is less than N0,
then ignore the company’s experience and use the external industry estimate as the forecast
value. Similarly NF is the sample size required for “full credibility”; if the company’s claim
costs are derived from a sample size that is larger than NF, then use company’s experience
and ignore the external industry estimate. Denote the number of policyholders in the most
recent company historical period by Ncomp, and the forecast value of PMPM claim cost by
Cforecast. According to credibility theory estimate Cforecast as follows:

• If Ncomp <= N0, then Cforecast = μind.

• If Ncomp >= NF, then Cforecast = μcomp.
• If Ncomp > N0 and Ncomp < NF, then Cforecast = λ x μcomp + (1 – λ) x μind.

Credibility Practice Note, July 2008

Page 9 of 55
In the actuarial literature listed in Section 5.1 below, there are two suggestions for the
determination of the weight, λ in the above formula. According to one suggestion - called
“limited fluctuation” - λ is determined as follows:

⎛ N comp ⎞
λ = min⎜ , 1⎟ .
⎜ NF ⎟
⎝ ⎠

According to another suggestion – called “greatest accuracy” - λ is determined as follows:

N comp
λ = .
N comp + K

In the above formula K is another parameter to be determined. In order to apply the above
formulas in practice the actuary needs estimates of the values of N0, NF, λ and K. There are,
however, no simple and generally accepted ways of estimating these parameters, and this fact
limits the practical usefulness of some credibility formulas.

DIFFICULTIES IN ESTIMATION OF N0, NF AND K

There are some problems encountered when we estimate the value of these parameters:

• N0 – Actuarial literature does not provide any method of estimating N0 that is

generally accepted and has a scientific basis.
• NF – The “limited fluctuation” approach provides the following formula for
derivation of NF:
z (21 − p ) σ 2
NF =
h2μ 2

However, to put the limited fluctuation credibility approach into practice, there are four
quantities that need to be determined: p, h, μ and σ2. Two of these quantities, namely, p and
h, are the confidence level and margin of error. By definition, each individual can choose
whatever value they want for these two quantities. The fact that it is a general industry
practice to work with, say, 95% confidence level and 5% margin of error does not sanctify
these two values and does give them “scientific respectability.” An actuary might want to
work with 90 percent confidence level - the p value - and five percent margin of error – the h
value; or on the other hand, they may be more comfortable working with a 99 percent
confidence level and one percent margin of error. The result is that for the same health
insurance policy, one actuary might consider the number “380” as a sample size that
produces fully credible experience data. Another actuary might consider any sample size less
than 1100 not sufficient to produce fully credible claims experience data.

Credibility Practice Note, July 2008

Page 10 of 55
The other two quantities, namely μ and σ2, are unknown. In fact μ is the unknown quantity,
namely the PMPM value, that we were trying to estimate in the first place. If the actuary
knew the value of μ there would be no need to estimate NF.

• K – In actuarial literature K is defined to be the ratio of ν/α, where ν is defined to

be “the expected process variance,” and α is defined to be “the variance of the
hypothetical means.” In general there are not good and acceptable estimates of
“the expected process variance” and “the variance of the hypothetical means.” In
fact, in general, there are no estimates of these parameters. Without good
estimates of these two quantities, estimation of K, and thereby, estimation of λ
becomes problematic.

APPLICATION OF CREDIBILITY THEORY TO HEALTH INSURANCE

When applying credibility theory to health insurance products the actuary should also take
into account the fact that policyholders that file for a claim stay as members of the group that
are exposed to the risk in future periods. This fact does not pose a significant problem when
the block of business under consideration has large number of policyholders. In that case the
fact that people filing for a claim remain as active members might have relatively small
effect on future PMPM claim costs. On the other hand, if a health insurance block has few
members, then the claim payment history of the members of the group becomes a crucial
piece of information. In other words, claims experience of a small – and closed - block of
health insurance business in one period is a good predictor of the claims experience in the
succeeding periods. The actuary doesn’t have such a problem in the case of life insurance
products. In that case, a policyholder who files for a claim, that is, if he or she dies, will no
longer be in the set of policyholders exposed to the mortality risk in future periods.

AN ECLECTIC USE OF THE THEORY OF CREDIBILITY IN PRACTICE

For the reasons mentioned above, it is not often, if at all, that we see a company actuary
using formula based credibility numbers to justify the company’s PMPM incurred claim cost
forecasts. Nor do the regulators require companies filing for rate increases to justify their
requests using some state-approved credibility formula. The instances where the departments
of insurance have sanctioned uses of some specific credibility factors are few. Does it mean
that departments of insurance ignore the concept of credibility? Not at all. Regulators look at
the company’s claims data, not just in the most recent year past, but for many years past, and
attempt to discover the extent to which the reported claims data reveal information regarding
the underlying claims experience of the company; the underlying claims experience that can
be used, together with many other sources of information, to predict the company’s claim
costs in future years. The important point to note is that in attempting to separate the noise
from the underlying trend, the regulators of health insurance products may choose not to use
“the limited fluctuation,” or “the greatest accuracy” approaches and formulas. Rather,
more often than not, regulators use educated common sense and judgment in their analysis of
company historical data.

Credibility Practice Note, July 2008

Page 11 of 55
3. Industry Practice

For some time, credibility theory has been applied within the property and casualty industry
in order to solve business problems. This has not been the case within the life, annuity and
health industries. Therefore, examples of the use of credibility theory and related practices
are somewhat difficult to find and somewhat simplistic in their content. With the advent of
principle-based approaches for the calculation of reserves and capital, there is a renewed
interest among actuaries concerning the use of credibility theory and related approaches in
the development of model assumptions.

The following examples are a mix of ways that credibility theory or related practices are
currently applied within the life, annuity or health industries and some possible solutions that
the authors would use to solve the stated problems. Therefore, they are not only intended to
illustrate some of the current practices but also some of the practices that have been
discussed in various actuarial groups. In order to help the reader become better acquainted
with the subject matter, a discussion of the strengths and weaknesses of a particular approach
used in the sample is provided after each example. Note that many of these examples
presume that an appropriate, identifiable external reference point exists for use with company
specific data. This point is also discussed in sections 2.2 and 3.4.

Credibility Practice Note, July 2008

Page 12 of 55
 3.1 A Limited Fluctuation Credibility Example

THE ISSUE

We want to estimate the mortality rate, q for a group of lives ages 50-59. We have one year’s
experience on each of 1,000 lives. For each life, j, we have recorded the amount of insurance,
bj. We have also recorded the outcome, dj = 1 if the life died, dj = 0 if the life lived. We
choose the dollar weighted estimate,

∑
1000
j =1 j
b dj Observed deaths ($)
qˆ = = .
∑
1000
bj Total insured ($)
j =1

What does it take for this estimate to be fully credible? Consider a limited fluctuation
approach and ask that the probability of the relative error being less than or equal to 5% be at
least 90%. These are arbitrary choices. Nothing in the limited fluctuation method helps set
these values. 1 Data are more likely to be fully credible if the 5% is increased or the 90%
lowered. To check for full credibility, calculate

⎛ | qˆ − q | ⎞
Pr ⎜ ≤ 0.05 ⎟
⎝ q ⎠

where q is the true probability and see if it is at least 0.9.

Three assumptions will be made:

• There are enough terms in the sum for the Central Limit Theorem to hold, 2
• The amounts of insurance are not random, and 3
• The lives are mutually independent and have the same value of q.

Then q̂ has a normal distribution with mean and variance:

∑ ∑
1000 1000
j =1
b j E (d j ) j =1 j
bq
E (qˆ ) = = =q
∑ ∑
1000 1000
j =1
bj b
j =1 j

∑ b Var (d ) = ∑ b q(1 − q) = σ
1000 2 1000 2
j =1 j j j =1 j
Var (qˆ ) = 2
.
(∑ b ) (∑ b )
1000 2 1000 2

j =1 j j =1 j

1
Over time, practice has typically set these values in the ranges 1-10% and 90-99%. If full flexibility is allowed, any
estimate can be declared fully credible or any partial credibility factor obtained.

Credibility Practice Note, July 2008

Page 13 of 55
Then,

⎛ | qˆ − q | ⎞ ⎛ −0.05q 0.05q ⎞
Pr ⎜ ≤ 0.05 ⎟ = Pr(−0.05q ≤ qˆ − q ≤ 0.05q ) = Pr ⎜ ≤Z ≤
⎝ q ⎠ ⎝ σ σ ⎟⎠

where Z has a standard normal distribution.

In order to perform this calculation, the unknown value of q must be estimated. The usual
approach is to substitute the estimate, q̂ . For example, suppose there were 200 policies with
a benefit of 10,000 and 3 of them died; 300 with a benefit of 25,000 and 7 died; 400 with a
benefit of 50,000 and 8 died; and 100 with a benefit of 100,000 and 3 died. The estimate is q̂
= 905,000/39,500,000 = 0.022911 and σ = 0.005628. The probability statement is then
Pr( −0.20355 ≤ Z ≤ 0.20355) = 0.1613. This is less than 0.9 and so the estimate is not fully
credible.

PARTIAL CREDIBILITY

With limited fluctuation credibility, when full credibility is not achieved, a weight, called the
partial credibility factor, must be determined. The most common method in use is the square
root rule. There are several justifications for this approach, all of them flawed. The rule
works as follows:

1. Determine the minimum exposure needed for full credibility.

2. The weight is the square root of the ratio of the actual exposure to the exposure
from step 1.

There is an easier way to get the same answer. Determine the standard normal value needed
to achieve the desired probability and take the ratio of the actual value to this value. For the
example, the partial credibility factor is 0.20355/1.645 = 0.1237.

This indicates that a weight of 12.37% should be given to the observed mortality probability
(relative frequency). Limited fluctuation credibility does not specify to what the remaining
weight should be applied.

SUMMARY OF LIMITED FLUCTUATION CREDIBILITY

Strengths:
• Good for experience rating, where there is a default premium.
• Simple to implement and understand.
• In certain cases, no estimation (only the number of claims) is needed to determine if
there is full credibility or to calculate the partial credibility factor.
Weaknesses:
• Reflects only the sampling variability of the data and not that of the base rate (the
quantity to which the remaining weight is applied).
• May not have an obvious base rate.

Credibility Practice Note, July 2008

Page 14 of 55
 • The choices of the two percents (5% and 90% in the example) are arbitrary.
• No sound statistical justification exists for the determination of the partial credibility
factor. In particular, most derivations use variance as a measure of quality. However,
credibility estimates are biased, making mean square error a more reasonable choice. 2

2
As an absurd example, the estimator “5” has a variance of zero, but is not a particularly good estimator of a
mortality rate.

Credibility Practice Note, July 2008

Page 15 of 55
 3.2 A Greatest Accuracy Credibility Example

Greatest accuracy credibility is the most commonly used alternative to limited fluctuation
credibility. It is also referred to as Bayesian credibility, linear Bayesian credibility, and
Buhlmann credibility (among other names).

All flavors of this method assume that there exists more than one entity (could be an infinite
number). Each entity produces numbers according to some probability distribution. For
example, the entity could be all the life insurance experience on one company’s block of
business and the number is the observed mortality ratio with respect to a standard table.

The second assumption is that these probability distributions are themselves distributed
among the entities according to another probability distribution. The goal is to then use this
information to estimate the probability distribution (or a key parameter of that distribution)
for one or all of the entities.

THE ISSUE

An illustration of this method is in the paper “Empirical Bayesian Credibility for Workers’
Compensation Classification Ratemaking” by Glenn Meyers (Proceedings of the CAS, Vol.
LXXI, 1984, pp. 96-121). I have slightly simplified the problem in this description. The
entities are 319 occupation classes for workers compensation insurance. 3 Each class
contributed three numbers, the average claim payments per covered worker in each of three
years. In addition, the number of covered workers in each class each year was recorded. The
objective is to estimate the true expected payments per worker in each class.

One possible model (not a very good one) is to make the following assumptions:

• All members of a given entity (occupation class) have the same distribution.
• Observations from a given entity both within one year and from year to year are
mutually independent.
• For an individual in a given class, the distribution of claims is normal with a mean
that depends on the class but a variance (v) that is the same for all classes and is
known.
• The class means are normally distributed with a known mean (μ) and variance (a).
• Mean square error will be minimized to determine the estimate.

This is not a good model in that the distributions are not likely to be normal. In theory, as
long as the two distributions (of claims and of class means) are completely specified, Bayes
Theorem can be used to obtain the solution. However, for many cases the computations can
be challenging. In the case of normal distributions there is no difficulty in obtaining the
Bayes solution. It is
n
Zx + (1 − Z ) μ , Z =
n+v/a

3
Sometime between 1984 and the present, the apostrophe was dropped.

Credibility Practice Note, July 2008

Page 16 of 55
 where n is the total sample size for the three years for the class and x is the sample mean for
a given class.

This approach has several challenges. First, the appropriate distributions must be identified.
Then, the parameters must be determined. Finally, execution of Bayes Theorem is not always
simple, depending on the distributions.

One way to ease the workload and reduce the number of assumptions is to insist in advance
that the answer take the form Zx + (1 − Z ) μ . Choose the answer that is closest (in the mean
square error sense) to the answer that would have been obtained had the true distributions
been known. It turns out that the answer, regardless of the true distribution, takes the form
given above. All that needs to be known are the moments μ, v, and a. This approach is often
called linear Bayesian, or Buhlmann credibility.

Finally, if values of the three moments are not known, they can be estimated from the data.
Formulas for doing this are in the Meyers paper and are also in many credibility textbook.
When this is done, the method is usually called empirical Bayesian credibility. One important
point – in order to estimate the parameters, there must be more than one observation from
each entity and data from more than one entity.

In all three cases (Bayes, linear Bayes and empirical Bayes), the formula makes sense. If
there are more observations (higher n), the sample mean gets more credibility (though never
full credibility). If v is high, then the observations are highly variable. This should imply that
the data are less credible, and that is what the formula indicates. The role of a is trickier.
Suppose a is large. Then the various entities are very different from each other. Thus any one
of them may be fairly large or small and thus should not be moved toward the middle. This
implies more credibility. On the other hand, if a were 0, every entity would have the same
true mean and there is no reason to give the data any credibility; every entity is average.
Another way to look at it is that a relates to the quality of μ in the same way that v relates to
the quality of the sample mean.

A MORTALITY EXAMPLE

Consider the previous example, but ignore the amounts. There were 1000 lives with 21
deaths. Suppose there was a second group with 2000 lives and 69 deaths. An empirical Bayes
approach would work as follows. 4

The mean, μ, is estimated via the sample mean over all observations, 90/3000 = 0.03.

The first variance is determined by estimating the variances for each group and then
weighting them by their sample sizes with a slight correction to make the estimator unbiased.
The result is

4
The formulas used here can be found on pages 595 and 596 of Loss Models: From Data to Decisions, 2nd ed by
Klugman, Panjer, and Willmot (Wiley, 2004).

Credibility Practice Note, July 2008

Page 17 of 55
 1000(0.021)(0.979) + 2000(0.0345)(0.9655)
vˆ = = 0.029079
999 + 1999

The estimate of a is less obvious, due to corrections needed to make it unbiased. It is

1000(0.021 − 0.03) 2 + 2000(0.0345 − 0.03) 2 − 0.029079(2 − 1)

aˆ = = 0.000069316
10002 + 20002
3000 −
3000

For the first group,

1000
Z= = 0.70447
0.029079
1000 +
0.000069316

and the credibility estimate is 0.70447(0.21) + 0.29553(0.03) = 0.023660.

SUMMARY OF GREATEST ACCURACY CREDIBILITY

Strengths:
• All aspects of the process are clearly spelled out. Any approximations or assumptions
made are clearly stated.
• Once the assumptions and objectives are in place, an exact answer can be obtained
from basic probability principles.
• There is a clear objective function – minimizing mean square error.
• There are no arbitrary choices that are unrelated to the observed random variables.

Weaknesses:
• The linear Bayes approximation may be poor, particularly if the random variable has
a heavy tail.
• The usual empirical Bayes estimate of a can be negative (because a is a variance, the
true value cannot be negative).
• Unless the prior distribution is known from some other knowledge source, it can only
be estimated if there is data from more than one entity.

Credibility Practice Note, July 2008

Page 18 of 55
3.3 Credibility for Mortality Ratios

This example shows how limited fluctuation credibility can be used for mortality ratios. For
each of i = 1,…,g ages or groups, you have collected ni observations. For the jth observation
from the ith group bij is the amount insured, fij is the fraction of year the person could have
been observed, and dij = 1 if the person dies. Also, qis is the standard table rate for group i
and you do not have qi which is the true rate. The estimated mortality ratio is

∑ ∑ ∑ ∑
g ni g ni
i =1 j =1 ij
b dij i =1
b dij
j =1 ij
mˆ = =
∑ ∑
g ni s
b f q e
i =1 j =1 ij ij i

where the denominator, e, is the known expected number of deaths. With no data we would
set m = 1 and use the standard table. This is where the remaining weight should be placed (a
common choice for limited fluctuation credibility is to ask what estimate would be used with
no data). In order to use limited fluctuation credibility we need the moments of the estimator:

∑ ∑ ∑ ∑
g ni g ni
i =1
b E (dij )
j =1 ij i =1
b f q
j =1 ij ij i
μ = E (mˆ ) = =
e e
∑ ∑ ∑ ∑
g ni g ni
i =1
2
j =1 ij
b Var (dij ) i =1
b f q (1 − f ij qi )
2
j =1 ij ij i
σ 2
= Var (mˆ ) = 2
= 2
.
e e

These can be best estimated by using the value mqˆ is for the unknown qi. This is because these
values will be more stable than using the estimates from the data. The formulas are:

∑ ∑
g ni
i =1 j =1 ij ij
ˆ is
b f mq
μˆ = = mˆ
e
∑ ∑
g ni
i =1
2
ˆ is (1 − f ij mq
b f mq
j =1 ij ij
ˆ is )
σˆ 2
= .
e2

As before, the key quantity is

rmˆ
zˆ =
σˆ

where r is the tolerance desired (0.05 in Example 3.1). Then, if ẑ is at least 1.645 (for 90%
confidence) the estimated mortality ratio can be given full credibility. If not, the estimate is

zˆ ⎛ zˆ ⎞
mˆ + ⎜ 1 − ⎟1 .
1.645 ⎝ 1.645 ⎠

Credibility Practice Note, July 2008

Page 19 of 55
The following data were supplied by Don Behan (some data adjustments were made for
simplification). The sample was 11,370 males from ages 70 through 100 who had purchased
charitable gift annuities. The data provided the number of life-years and the actual and
expected number of deaths at each age. The data already took into account partial-year
exposure, so the variable f in the above formulas is not needed. Because the mortality rates
were based on lives, rather than on benefit amounts, the value of b in the formulas was set to
1. There were 782.67 expected deaths (US 2000 Annuity Table) and 744 actual deaths for an
estimated mortality ratio of mˆ = 0.9506 .

The variance estimate is 0.0011. Set the full credibility standard as being within 5% of the
true ratio 95% of the time. The standardized variable is zˆ = 0.05(0.9506) / 0.0011 = 1.432 .
This is below 1.96, and therefore full credibility cannot be granted.

The partial credibility factor is 1.432/1.96 = 0.7306 and then the credibility estimate for the
mortality ratio is 0.7306 x (0.9506) + (1 - .7306) x 1 = 0.9638.

The following conclusions and question can be drawn from applying this method to
estimated mortality ratios:

• This method is easy to use. It is likely the data required are available and if not, a
reasonable approximation may be available.
• It is not terribly arbitrary (as long as bounds on the tolerance and probability are
agreed upon).
• The pros and cons of limited fluctuation credibility apply here.
• Why is 1 the number being multiplied by the complement of credibility? (Because we
are assuming that the US 2000 Annuity Table would be used in the absence of
credible data for this particular group.)

OTHER OPTIONS FOR MORTALITY RATIOS

If we had data from more than one group we could then use greatest accuracy credibility.
One immediate advantage is we may learn that the industry average is not the standard table,
but some multiple of it.

On the other hand, this may not be a credibility problem at all. We are trying to estimate only
one thing. We do not care if the error is reduced over all companies, only the error for our
company. Won’t there be times when 1 is not the right starting point? Maybe our block is
better (or worse) for reasons that are independent of the data (marketing, underwriting
distinctions, etc.)

Instead, start with this question: What mortality ratio is used when data are limited? Consider
a two-step process?
1
• If m̂ ± 1.96σˆ includes 1, use the standard table. This is equivalent to doing a
hypothesis test. If you cannot reject the hypothesis that the mortality ratio is 1,
consider using the standard table.

Credibility Practice Note, July 2008

Page 20 of 55
 • If the interval does not contain 1, use one of the endpoints of the interval. The
appropriate endpoint depends on the ultimate use of the ratio. The interval will not
contain 1 either if the estimated ratio is far from 1 or if the sample size is large
(reducing the standard deviation). Either case is an argument for not using the
standard table. However, in all cases, estimation error will still be accounted for (so
there is never full credibility).

In the example, 1 is in the confidence interval, so this method would indicate that the data do
not justify deviating from the standard table.

Credibility Practice Note, July 2008

Page 21 of 55
3.4 Practical Considerations for Credibility in Reinsurance Pricing

In pricing individual life reinsurance for a client (a direct insurer), an application of

credibility theory is needed to blend a starting-point mortality table with the client’s own
mortality experience. In many cases the starting-point mortality table is the reinsurer’s
client-specific mortality table death rates derived from seriatim internal and client-provided
data, with adjustments to take into account underwriting requirements and preferred criteria,
but set to an overall level of mortality that may differ from experience for the client whose
account is being priced. One of the most important questions is typically, “how closely
should the pricing mortality resemble the client’s experience?” The first step in determining
an answer to this question is to apply credibility theory, but there are several important
matters to consider before finalizing the mortality assumption.

One major consideration is how relevant the client’s mortality experience is to the product
being priced. If the product specifications, underwriting requirements, preferred criteria,
underwriting staff, and distribution channels are all identical between the product being
priced and the available mortality experience, then this should not be an issue. But if a new
chief underwriter has since taken over, requirements have been strengthened, and a preferred
class has been added, then perhaps a little less emphasis should be placed on the experience,
and the effect of these changes should be incorporated into the pricing mortality by means
other than the application of credibility.

Another consideration is the base mortality table that is used. If the experience is being
blended with a mortality table that is very credible but is generic to the insurance industry
and is based on experience from an era when not much preferred business was being issued,
then the pricing mortality could be allowed to be more heavily influenced by a small to
medium amount of credibility in the client’s experience. But if the experience is being
blended with a credible mortality table that was derived from recent insured lives experience
from other companies but adjusted to account for the client’s own underwriting, claims, and
marketing practices, then it might take more experience data from the client to have as much
of an effect on the pricing mortality.

Yet another consideration is to what extent the mortality slopes (relative mortality by age,
gender, duration, risk class, etc.) in the client’s experience should be reflected in the pricing
mortality. In most situations, the credibility of the overall experience is questionable enough
that partitioning the experience data to compare mortality among subgroups results in such
low credibility that it is not worth the effort of going through the exercise of blending the
experience with the base mortality table at the cell level. But situations do arise when there
is a different mortality pattern that is specific to a certain company and there is enough
credibility in the subgroups that it should be incorporated into the pricing mortality. When
this happens, credibility theory can be applied to determine to what extent that pattern is
reflected in the mortality assumption and the overall level of mortality can be further adjusted
based on the credibility of the experience in total.

Credibility Practice Note, July 2008

Page 22 of 55
Once all of these considerations have been addressed, the pricing actuary can determine
adjustments to make to the credibility weight that is to be applied to the client’s mortality
experience before blending it with a mortality table.

To illustrate an application of credibility in reinsurance pricing using an example, consider a

situation where an actuary is pricing a new universal life product for ABC Life. ABC Life
has only issued term policies in the past, so they supplied experience data for those plans,
since it will be sold through the same distribution channels and the underwriting staff is the
same. However, with the introduction of the new UL product, they are adding a super
preferred non-tobacco class.

ABC Life’s mortality experience study results in an overall actual-to-expected (A/E) ratio of
110% relative to the reinsurer’s base mortality table. The pricing actuary notices unusually
high mortality in the preferred tobacco class, but since there are only 20 claims in that class
the actuary decides to use the relationships already established in the base table rather than
go through the effort to reshape the mortality slope by preferred class for this client with this
limited amount of additional experience. Instead, the actuary will focus efforts on using the
client’s experience to make an adjustment to the overall level of mortality in the base table.
More discussion on addressing client-specific mortality slopes appears later in this example.

Note that the A/E ratio in the experience study is relative to a version of the base table that
only has two non-tobacco classes, while the base table used for pricing the new product was
derived on an equivalent basis, but includes the additional super preferred class.

For illustrative purposes in this example, let’s say the actuary, using formulas and parameters
consistent with prior reinsurance pricing exercises, arrives at a credibility weight of 70%.

Since the study did not include a super preferred class, the actuary may choose to reduce the
credibility factor by several percent, perhaps to 63%. The study being based on term and not
UL might suggest that the percentage be further reduced to 60%. It should be noted that in
this example these reductions to the credibility weight are based strictly on actuarial
judgment.

A credibility weight of 60% and an A/E ratio of 110% results in the actuary setting the final
pricing mortality assumption at 106% ( = 60% x 110% + 40% x 100%) of the base table.

Getting back to client-specific mortality slopes, it should be noted that if there is an

underlying reason for a client’s experience to show mortality by class within any given
category that has a different shape than what the base table assumes, then it should probably
be reflected in the mortality assumption used in pricing that client’s reinsurance. In the
above example, if there is no reason to believe the elevated mortality in the preferred tobacco
class is caused by something systematically different in that client’s underwriting relative to
all other clients, then it could be assumed that it was just a random fluctuation since there
were only 20 claims. It should not be reflected in the mortality assumption by adjusting
preferred tobacco mortality up and the other classes down.

Credibility Practice Note, July 2008

Page 23 of 55
An example where an actuary should consider a client-specific slope adjustment might be
where a client includes comments in their underwriters’ training manual to direct them to
allow fewer exceptions when applying preferred criteria for ages 50 and above, and as a
result their mortality experience shows lower A/E ratios at these ages. If no adjustment is
made to the age slope in the mortality table, the reinsurer is left open to distribution risk,
meaning that if the age distribution of policies issued in the future shifts toward the younger
ages, the overall A/E ratios of future experience studies will go up. Also, if the client
insurance company does not reflect in the premiums the differences by age in how the
company applies the preferred criteria, it will likely cause premiums for ages 50 and above to
be less attractive to potential applicants, speeding up the undesirable shift in the age
distribution.

Credibility Practice Note, July 2008

Page 24 of 55
3.5 Practical Considerations for using Credibility for Mortality

This example describes practical considerations in applying credibility theory to develop or

support mortality assumptions. The example presents a practical framework in approaching
the setting of mortality assumptions. The example uses an approach that is an admittedly less
theoretically grounded approach than others in practice.

MORTALITY ASSUMPTION

When actuaries build their models, they use assumptions they believe appropriate. For
mortality, this is usually k% or .01K of a known mortality table.

PRE STUDY DATA HANDLING

Inforce files and transaction data are checked and reconciled and any data deficiency is
corrected or noted before an experience study is completed.

MORTALITY STUDY

There are factors to consider when performing a mortality study. It could be a valuation study
which uses grouped inforce data at valuation dates and the decrement transactions in between
valuation dates. It could be a seriatim study which tracks the status of each policy thru a
study period.

Some mortality studies aim to create a new mortality table, using the actual number or
amount of deaths/exposure as an estimate of the death rate.

Another fairly common approach is to compute an expected number or amount of deaths

based on a specific table by multiplying the exposure amount by the known tabular death
rate. This expected amount may also be known as the tabular amount or number of deaths,
since it is the tabular death rate that was applied to the exposure. The actual deaths-to-tabular
deaths (A/T) ratio is the experienced tracked. Another name for this A/T Ratio is the
Mortality Ratio (MR). For this example, this is the mortality study approach adopted.

Some considerations in the grouping of the results are product or death benefit type, age
brackets, durations or issue year groupings. These considerations are to be consistent with the
way the assumptions will be used in the actuarial models.

USE OF CREDIBILITY

Credibility may be used when the mortality experience is to be weighted against either a
prescribed mortality table or a known industry standard table.

A simple approach is to weigh the experienced MR by credibility factors based on the

number of deaths. The company may either decide to develop its own credibility weights or
use factors developed in existing literature.

Credibility Practice Note, July 2008

Page 25 of 55
One set of credibility factors, can be found in the CIA Education Note on Expected Mortality
completed in July 2002. In Section 540 and Appendix 2 of that note are descriptions of
Limited Fluctuation Credibility Theory and the development of the number of claims in order
to have full credibility. In the same report, the development of partial credibility factors is
presented and the choice of 3007 claims as a threshold for z = 1 (full credibility), was arrived
at by consensus. Descriptions of Limited Fluctuation Credibility, its limitations, and
determining the claims to meet full credibility are discussed in many of the credibility books
listed in Section 5.1. One such book is Introduction to Credibility Theory by Thomas Herzog.
Page 60 of that book, shows how full credibility may be determined using minimum number
of expected claims. In practice, the actual number of claims is used.

As an example, in that cited CIA Education Note the following credibility factor table was
developed for use in blending inter-company table and company experience:

90% Probability of being correct

within 3% margin of error

Number of claims Credibility factor Z

30 .10
120 .20
271 .30
481 .40
752 .50
1,083 .60
1,473 .70
1,924 .80
2,436 .90
3,007 1.00

Full Credibility Z = 1 is assumed to occur at 3007 claims, while

Partial Credibility Z = Minimum {(number of claims / 3007) ^.5, 1}

The credibility weighted MR =

Z * Experienced MR + (1 – Z) * prescribed or assumed MR

The credibility weighted MR would be compared to the assumed MR.

Should the actuary decide to investigate the credibility at each of the mortality curves
segments, the same CIA report has a Section 540 entitled Normalized Method of Limited
Fluctuation Credibility Theory, which shows a method of normalizing the weighted
experience at each segment such that overall weighted experience is preserved.

Credibility Practice Note, July 2008

Page 26 of 55
NUMERIC EXAMPLE

The Actuaries at ABC Insurance Company have set their mortality assumption as [or ‘at’?]
60% of the XYZ mortality table. This is the assumption they use for all the financial
projections and pricing work. The Experience Studies Actuary has the task of monitoring the
experience of the company and is charged with proposing changes to the company’s current
assumptions.

Some preliminary testing has led the Experience Study Actuary to conclude that ABC’s
business belongs to the plus segment. “Plus segment” is a term used in RBC’s C3P2 to
describe a situation whereby a given financial measure deteriorates as mortality increases.
This gives the actuary the conservative sense of direction, i.e., a higher mortality assumption
than indicated by experience would be on the conservative side.

After the necessary data checking is done, Table 1 shows the result of the study with
computed confidence interval.

Table 1: Mortality Ratio Confidence Interval

ABC Mortality Study

Study YYYY and YYYY+1
Mortality Ratios
Based on XYZ Mortality Table
Confidence Intervals

(1) (2) (3) (4) (5) (6) (7) (8)

95% Conf Interval
AcctVal Deaths (000's) Actual / Standard Bounds of A/T
Age # Expected Tabular
Group deaths Tabular Actual (A/T) Deviation Lower Upper

00 - 50 69 15,713 6,932 44.12% 5.31% 33.71% 54.53%

51 - 70 443 180,477 67,840 37.59% 1.79% 34.09% 41.09%
71+ 835 281,162 149,044 53.01% 1.83% 49.41% 56.61%

total 1,347 477,352 223,816 46.89% 1.28% 44.38% 49.39%

It can be seen that the experience is less than the assumed 60% of XYZ Mortality. Even the
upper bound of the experience is less than the assumed 60% of XYZ mortality. These are
true regardless of whether viewed from the whole mortality curve or each of the three
segments. It appears that the assumed mortality has some margins. The actuary may decide to
keep this margin.

The actuary would also want to credibility weight the experience against an industry
standard. To do this, he/she has to determine three issues: (1) the credibility weight formula
to use; (2) the number of deaths to give full credibility to; and (3) the industry standard table
to counter weight the experience.

Credibility Practice Note, July 2008

Page 27 of 55
The Experience Study Actuary has decided to use the following credibility weight formula:

Z = minimum { (number of deaths / number of deaths with full credibility)^.5, 1) with the
full credibility assumed to occur at 3007 deaths.

The actuary has based this on Section 540 and Appendix 2 of the July 2002 CIA Education
Note. In Section 540, the consensus minimum number of deaths recommended for 100%
credibility was 3007. A more theoretical approach would have been to actually compute the
credibility weights as described elsewhere in the practice note.

The Experience Study Actuary chose the 1994 MGDB mortality table as the standard table to
“counterbalance” against the experience. This was chosen on three grounds: (1) it is a
recognizable standard table; (2) it was based on an industry wide study; and (3) it is the
prescribed mortality table for similar products.

Since the company uses the XYZ table which is different from the standard table (e.g., 1994
MGDB), it would be necessary to translate the standard table to a percentage of the XYZ
table that the company is using. It was determined beforehand that the XYZ mortality for this
business is 133% of the standard valuation table (e.g., the 1994 MGDB mortality table is
75% of XYZ mortality). For each of the age segments, the expected based on the standard
valuation table to the expected based on the tabular XYZ mortality table were noted and will
be used as the “industry” actual-to-expected ratio counterbalancing the experience. To
illustrate this, in Table 2 the ratio of the Standard table to table that the company uses is
42.96% for ages 0-50, 70% for ages 51-70, 80% ages 71 above, for an overall ratio of the
Standard table of 70% of the Table the company uses.

Table 2 shows the credibility weighting of the mortality experience (overall 66.93% of XYZ
table) versus that expected by the standard industry table (overall 75% of XYZ table).

Table 2: Credibility Weighted Mortality Ratio

ABC Mortality Study

Study YYYY and YYYY+1
Mortality Ratios
Based on XYZ Mortality Table
Credibility Weighted

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
AcctVal Deaths (000's) Actual/ weight Z Standard Standard Cred.Wtd Expected Deaths
Age # Expected Tabular Table Table
Group deaths Tabular Actual (A/T) Expected to Tabular Mort Ratios Credibility Weighted

00 - 50 69 15,713 6,932 44.12% 15.15% 6,751 42.96% 43.14% 6,778 39.14%

51 - 70 443 180,477 67,840 37.59% 38.38% 126,334 70.00% 57.56% 103,882 52.22%
71+ 835 281,162 149,044 53.01% 52.70% 224,930 80.00% 65.78% 184,941 59.68%

total 1,347 477,352 223,816 46.89% 66.93% 358,014 75.00% 56.18% 268,196 56.18%
295,601
Normalization
factor: 0.90729

Column (5) is the Actual to Tabular Mortality Ratio Raw experience.

Credibility Practice Note, July 2008

Page 28 of 55
Column (6) is the credibility weights to be used based on Full Credibility.
Column (8) is the industry Mortality Ratio to be counter weighted against.
Column (9) is the credibility weighted Mortality Ratio experience.

The overall credibility mortality ratio is 56.18% which is less than the assumed 60%.

The actuary may normalize the MR at each age segment so that the sum of expected deaths
for all the segments preserves the total produced by the blended overall MR. This may be
seen in Column (11).

The actuary may observe the blended mortality ratio at the overall level (56.18%) and at the
segment level after normalization (39.14%, 52.22%, 59.68%) and draw a conclusion as to
acceptability of keeping the current mortality ratio assumption or redo the MR assumption.

Knowing that this block of business belongs to the plus segment gives the actuary a sense of
direction as far as whether margins exist. In this case, experience is favorable from both the
confidence interval view (upper bound of the experience is less than assumed MR) and the
overall credibility weighted MR (lower then the assumed MR). This would lead the actuary
to propose no change in mortality assumption for this block.

The experience for the block is monitored each year and any historical trends are considered
in the decision to keep or revise the assumed mortality for the block.

Credibility Practice Note, July 2008

Page 29 of 55
3.6 An Example Using Limited Fluctuation Credibility to Update Lapse Assumptions

An Insurance company is undergoing the process of updating the lapse assumptions in their
projection models. The same process has been used for several years. At a high level, the
process requires three basic steps: the first is to obtain experience data from the Insurance In-
force Management group; the next step is to enter the data into an existing worksheet; the
final step is to review the results and update the projection model. The experience data is for
a Variable Universal Life (VUL) product with seven study-years (1999-2005)
Actual Lapse Rates= (LAPSE+REPLACED+SURRENDER)/(LAPSE+REPLACED+SURRENDER+INFORCE)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16+
0 to 34 7.0% 10.6% 9.5% 8.7% 7.5% 8.4% 7.7% 7.3% 7.7% 7.3% 8.4% 7.6% 6.8% 6.6% 6.7% 7.2%
35 to 44 6.2% 9.1% 9.0% 7.5% 6.9% 7.0% 6.8% 6.3% 6.1% 6.5% 7.9% 7.0% 7.2% 7.1% 7.6% 7.5%
45 to 54 5.1% 7.4% 7.2% 6.6% 5.8% 6.3% 6.3% 6.1% 6.5% 7.2% 8.6% 8.7% 7.7% 8.9% 8.4% 7.7%
55 to 59 4.1% 6.2% 5.0% 5.8% 5.5% 6.1% 6.5% 6.2% 6.4% 8.8% 10.8% 11.4% 8.5% 11.7% 11.1% 10.8%
60 to 64 2.7% 5.4% 4.2% 6.1% 5.5% 6.2% 6.5% 7.1% 6.5% 8.7% 10.6% 10.2% 10.7% 12.4% 9.4% 12.4%
65 to 69 3.8% 4.8% 6.2% 6.5% 6.6% 7.3% 8.5% 6.1% 8.6% 8.2% 8.1% 8.9% 6.6% 11.2% 2.9% 12.3%
70 to 74 2.3% 4.5% 6.1% 6.8% 8.1% 7.8% 9.8% 6.7% 10.8% 8.7% 10.8% 9.5% 8.9% 17.8% 7.4% 3.3%
75 to 99 1.9% 9.5% 6.4% 8.9% 6.1% 9.7% 10.3% 10.0% 8.0% 7.1% 10.7% 2.1% 10.5% 4.9% 1.8% 10.1%

Exposure Count Policies = n

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16+
0 to 34 25,090 30,887 35,307 39,154 37,684 36,282 33,285 30,283 26,337 22,554 18,568 16,279 12,927 9,387 6,556 10,659
35 to 44 18,695 24,447 29,859 35,118 35,928 36,187 33,975 30,878 27,047 22,716 17,785 14,959 11,338 7,926 5,482 8,523
45 to 54 14,124 18,099 21,801 25,641 25,352 24,571 21,832 18,834 15,688 12,329 8,626 6,866 4,756 3,134 2,099 3,170
55 to 59 3,627 4,665 5,748 6,810 6,612 6,355 5,587 4,717 3,873 2,930 1,983 1,619 1,114 747 511 767
60 to 64 1,809 2,446 3,137 3,802 3,772 3,650 3,260 2,785 2,274 1,715 1,129 899 617 409 277 405
65 to 69 988 1,373 1,777 2,153 2,177 2,137 1,924 1,591 1,306 978 663 523 333 223 152 222
70 to 74 562 868 1,155 1,418 1,409 1,368 1,227 1,016 808 582 372 281 195 130 88 107
75 to 99 436 618 772 945 902 847 713 564 425 300 180 127 84 52 39 52

The process consists of 4 steps:

1) The table below is in the worksheet. This is the “base” table for credibility application.
Assumed Base plan lapse rate = p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16+
0 to 34 3.0% 3.0% 5.0% 7.0% 7.0% 7.0% 7.0% 7.0% 7.0% 7.0% 7.0% 7.0% 7.0% 7.0% 7.5% 8.0%
35 to 44 2.0% 2.0% 3.5% 4.0% 4.5% 4.5% 4.5% 5.0% 5.5% 5.5% 6.5% 6.0% 6.5% 7.0% 7.5% 8.0%
45 to 54 2.0% 2.0% 3.0% 3.5% 4.0% 4.0% 4.5% 5.0% 5.5% 5.5% 6.5% 6.0% 6.5% 7.0% 7.5% 8.0%
55 to 59 2.0% 2.0% 3.0% 3.5% 4.0% 4.0% 4.5% 5.0% 5.5% 5.5% 6.5% 6.0% 6.5% 7.0% 7.5% 8.0%
60 to 64 2.0% 2.0% 3.0% 3.5% 4.0% 4.0% 4.5% 5.0% 5.5% 5.5% 6.5% 6.0% 6.5% 7.0% 7.5% 8.0%
65 to 69 2.0% 2.0% 3.0% 3.5% 4.0% 4.0% 4.5% 5.0% 5.5% 5.5% 6.5% 6.0% 6.5% 7.0% 7.5% 8.0%
70 to 74 2.0% 2.0% 3.0% 3.5% 4.0% 4.0% 4.5% 5.0% 5.5% 5.5% 6.5% 6.0% 6.5% 7.0% 7.5% 8.0%
75 to 99 2.0% 2.0% 3.0% 3.5% 4.0% 4.0% 4.5% 5.0% 5.5% 5.5% 6.5% 6.0% 6.5% 7.0% 7.5% 8.0%

2) Next, the Expected Lapse Count is calculated by multiplying the Assumed Base Plan
Lapse rate by the Exposure Count.
Expected Lapse Count =n*p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16+
0 to 34 753 927 1,765 2,741 2,638 2,540 2,330 2,120 1,844 1,579 1,300 1,140 905 657 492 853
35 to 44 374 489 1,045 1,405 1,617 1,628 1,529 1,544 1,488 1,249 1,156 898 737 555 411 682
45 to 54 282 362 654 897 1,014 983 982 942 863 678 561 412 309 219 157 254
55 to 59 73 93 172 238 264 254 251 236 213 161 129 97 72 52 38 61
60 to 64 36 49 94 133 151 146 147 139 125 94 73 54 40 29 21 32
65 to 69 20 27 53 75 87 85 87 80 72 54 43 31 22 16 11 18
70 to 74 11 17 35 50 56 55 55 51 44 32 24 17 13 9 7 9
75 to 99 9 12 23 33 36 34 32 28 23 17 12 8 5 4 3 4

Credibility Practice Note, July 2008

Page 30 of 55
3) Next a credibility factor, Z, is calculated as

⎛ Expected Lapse Count ⎞

min⎜⎜1, ⎟
1,082 ⎟
⎝ ⎠

Credibility Factor = Z = √(Expected Lapse Count/1082)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16+
0 to 34 0.83 0.93 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.91 0.78 0.67 0.89
35 to 44 0.59 0.67 0.98 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.91 0.83 0.72 0.62 0.79
45 to 54 0.51 0.58 0.78 0.91 0.97 0.95 0.95 0.93 0.89 0.79 0.72 0.62 0.53 0.45 0.38 0.48
55 to 59 0.26 0.29 0.40 0.47 0.49 0.48 0.48 0.47 0.44 0.39 0.35 0.30 0.26 0.22 0.19 0.24
60 to 64 0.18 0.21 0.29 0.35 0.37 0.37 0.37 0.36 0.34 0.30 0.26 0.22 0.19 0.16 0.14 0.17
65 to 69 0.14 0.16 0.22 0.26 0.28 0.28 0.28 0.27 0.26 0.22 0.20 0.17 0.14 0.12 0.10 0.13
70 to 74 0.10 0.13 0.18 0.21 0.23 0.22 0.23 0.22 0.20 0.17 0.15 0.12 0.11 0.09 0.08 0.09
75 to 99 0.09 0.11 0.15 0.17 0.18 0.18 0.17 0.16 0.15 0.12 0.10 0.08 0.07 0.06 0.05 0.06

4) Finally, the revised lapse rates are computed as

Z ⋅ ( Experience Lapse Rate ) + (1 − Z) ⋅ p

Credibility Adjusted Factors: Z x (experience lapse rate) + (1-Z) x p

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16+
0 to 34 6.3% 10.1% 9.5% 8.7% 7.5% 8.4% 7.7% 7.3% 7.7% 7.3% 8.4% 7.6% 6.8% 6.7% 7.0% 7.3%
35 to 44 4.5% 6.8% 8.9% 7.5% 6.9% 7.0% 6.8% 6.3% 6.1% 6.5% 7.9% 6.9% 7.1% 7.1% 7.6% 7.6%
45 to 54 3.6% 5.1% 6.2% 6.3% 5.8% 6.2% 6.2% 6.0% 6.4% 6.8% 8.0% 7.7% 7.1% 7.9% 7.8% 7.9%
55 to 59 2.5% 3.2% 3.8% 4.6% 4.7% 5.0% 5.5% 5.5% 5.9% 6.8% 8.0% 7.6% 7.0% 8.0% 8.2% 8.7%
60 to 64 2.1% 2.7% 3.4% 4.4% 4.6% 4.8% 5.2% 5.8% 5.8% 6.5% 7.6% 6.9% 7.3% 7.9% 7.8% 8.8%
65 to 69 2.2% 2.4% 3.7% 4.3% 4.7% 4.9% 5.6% 5.3% 6.3% 6.1% 6.8% 6.5% 6.5% 7.5% 7.0% 8.5%
70 to 74 2.0% 2.3% 3.6% 4.2% 4.9% 4.9% 5.7% 5.4% 6.6% 6.1% 7.1% 6.4% 6.8% 8.0% 7.5% 7.6%
75 to 99 2.0% 2.8% 3.5% 4.4% 4.4% 5.0% 5.5% 5.8% 5.9% 5.7% 6.9% 5.7% 6.8% 6.9% 7.2% 8.1%

While the process is straight forward, it is worth mentioning key strengths and weaknesses.

Strengths:

• The method is relatively simple, with the desired result of bringing the company’s
experience into the new assumption.
• Cells with over 1082 exposures are considered to have full credibility.
• The method is easily explainable and auditable.
• The level of granularity allows for variation among cells to impact the final assumptions.
Weaknesses:

• Limited-Fluctuation Credibility has no sound theoretical basis.

• The choice of the Assumed Base Plan Lapse table needs to be justified.
• Using 1082 as the denominator in determining Z needs to be justified. Using a different
basis for full credibility could have a large impact on the results.
• The level of granularity of the data needs to be justified; aggregation among the cells
would lead to full credibility more quickly.

Credibility Practice Note, July 2008

Page 31 of 55
• The actuary needs to understand the experience data and determine whether it is
appropriate to use for the future.
• With the above choices, different actuaries may get different results using the same data.
• The results may not be “smooth.”

ADDITIONAL CONSIDERATION

There are alternative approaches that an actuary could use to build assumptions from
experience data. Two examples are an application of regression or Bayesian credibility.

Credibility Practice Note, July 2008

Page 32 of 55
3.7 Credibility Conflict I

PROBLEM

A credit insurer has never had any insurance claims but refuses to lower its premium rates.
The state’s chief actuary thinks that this is causing insureds to be overcharged.

Thus far the insurer has insured one million policies without a claim. How many claims
should be expected on its next one million policies?

Note: This example is an extreme case and is used to make a point rather than to illustrate a
situation likely to be encountered in practice.

SOLUTION

General Approach - We adopt a Bayesian approach.

Assumptions - We make the following assumptions:

First, we assume that claims are mutually independent and identically distributed with a
binomial distribution with parameter, Θ . So, we use the binomial with parameter, Θ , as our
likelihood.

Second, we assume that the parameter, Θ , has a prior beta distribution, B(a,b), with
parameters a = 1 and b = 999 , so that the (prior) mean of the beta distribution is
a 1
= = .001 .
a + b 1 + 999

This estimated mean was obtained from the experience of a large number of policies of some
related insurance. This conforms to the approach advocated by Buhlmann as discussed in
Section 4 of this work. The implicit sample size selected here is

a + b = 1 + 999 = 1,000 .

Some practitioners might feel that we should have used a larger implicit sample size;
perhaps, we should have chosen a = 100 and b = 99,900 .

Note: Here we are treating Θ as a random variable.

Computations

Following the discussion of (1) Theorem 8.1 on pages 135-136 and (2) Exercise 8-2(a) of
Herzog [1999], we find that the posterior distribution of Θ is B(a+r,b+n-r) where r = 0 is

Credibility Practice Note, July 2008

Page 33 of 55
the number of observed “successes” – i.e., claims – and n = 1,000,000 is the number of trials
– i.e., exposures, measured by policies. The mean of this posterior distribution is

a+r 1+ 0 1
= = .
a + b + n 1 + 999 + 1,000,000 1,001,000

This is a substantial decline from the prior mean that we began with.
So, the expected number of claims on the next 1,000,000 policies is

1
(1,000,000) ⋅ ≅ 1.
1,001,000

Strengths:

• This approach conforms to the underlying mathematical theory.

• All of the assumptions made are set forth explicitly.

• The result obtained is a reasonable one. This is in sharp contrast to the results of a
limited fluctuation approach to this problem. Under that paradigm, you end up with a
credibility factor of either Z = 1 or Z = 0. For the first case, you end up giving all of the
weight to the current observations and so the probability of any future claims is zero – a
result that is too low (you want a small positive quantity). For the second case, you end
up giving all of the weight to the prior data and so you end up with a probability of .001
of a claim on an individual trial or 1,000,000 ⋅ (.001) = 1,000 expected claims on the next
1 million policies – a result that is almost certainly too high.

Weaknesses:

• All models require the actuary to make assumptions. So, some actuaries might prefer to
employ a different type of prior distribution and/or to use a different set of prior
parameters.

Credibility Practice Note, July 2008

Page 34 of 55
3.8 Credibility Conflict II

A company enters the state credit insurance market. The state requires that rates be set at a
level to produce a 50% loss ratio. The state also publishes a “prima facie” rate that is
presumed to produce 50% loss ratio. Let’s say that this rate is $8 per $1,000 of coverage per
year. That corresponds to an incidence rate of 4 claims per 1,000 policies per year.

The state requires an annual experience report to be filed. Past experience may indicate a
need for deviation from the prima facie rate. Because the number of claims per year is small
and there is little variation in policy size, experience is reported as claim counts rather than
claim dollars.

In the first year the company issues 5,000 policies (all on the first day of the year) and has 8
claims, for a claim rate of 0.0016 which is well below the prima facie rate of 0.0040. As a
result, the regulatory actuary is considering requesting that the company reduce its premium
rate.

REGULATOR’S CASE

The regulator takes a Bayesian approach. Let X denote the number of claims for the
company in a given year. It is reasonable to assume that X has a binomial distribution with n
= 5,000 and p unknown. The prima facie rate assumes p = 0.004, but individual companies
may have different true values of p.

In order to use the Bayesian approach, a prior distribution is required. There are two
equivalent (in the sense that they lead to the same formulas) interpretations of this
distribution. One is that it reflects the regulator’s confidence in the prima facie value of
0.004. The other is that it reflects how the parameter p varies from company to company.
The regulator selects a beta distribution with parameters a = 16 and b = 3,984. The beta
distribution is selected for two reasons. First, it is easy to work with; all the calculations
needed are simple to execute. Second, it has an appropriate shape. Below is a graph of this
density function.

The parameters were selected so that the mean is the prima facie value of 0.004. The
standard deviation of this beta distribution is about 0.001 and this also provides an
interpretation of his choice.

Credibility Practice Note, July 2008

Page 35 of 55
 p(p)
450
400
350
300
250
200
150
100
50
0
0 0.002 0.004 0.006 0.008 0.01

The key output of a Bayesian analysis is the posterior distribution. This represents the
regulator’s opinion about the distribution of p for the company in question based on
combining the data from the company with the prior distribution. It turns out that the
posterior distribution is also beta. The parameters of this new distribution are a*= 16 + 8 =
24 and b* = 3,984 + 4,992 = 8,976 (this outcome is derived in all credibility texts). The
posterior distribution is plotted below. It can be seen that the spread is reduced (indicating
increased confidence) and that the mode has moved to a smaller value (reflecting the
company experience).

p(p)
800
700
600
500
400
300
200
100
0
0 0.002 0.004 0.006 0.008 0.01

The mean, 24/9,000 = 0.00267 is the Bayesian credibility estimate. This can be interpreted
as applying a weight of 5/9 to the company’s experience and 4/9 to the prima facie rate.

The premium based on this result is 2(0.00267)(1000) = $5.34 per thousand, a 33% decrease.
This is the regulator’s recommendation (in the above formula is the multiplier ‘2’ is obtained
by dividing by the loss ratio of 50%).

Credibility Practice Note, July 2008

Page 36 of 55
COMPANY’S CASE

The company actuary vehemently objects on the basis that the data lacks credibility. He
demonstrates this using limited fluctuation credibility. This method also appears in every
credibility textbook and has been used for decades with a wide variety of products.

The key assumption is that the claims distribution can be approximated with a normal
distribution. Normally, this requires at least 5 expected claims. Under the prima facie
assumption, with 5,000 policies, there are 20 claims expected. The next step is to set a
standard for full credibility. The company has set the standard as the minimum number of
policies so that the estimated probability of a claim has a relative error of 5% with 90%
confidence. When the mean and variance are equal (in the prima facie case the mean is 5 and
the variance is 4.98) the standard for full credibility is 1,083 claims. When there are fewer
claims, a square root rule is employed to determine the credibility factor.

In this case, with 8 claims, the credibility factor is the square root of 8/1,083 = 0.08595 and
thus the appropriate estimate is 0.08595(0.0016) + 0.91405(0.004) = 0.003794 which implies
a premium of $7.59. This translates to a 5% premium reduction.

QUESTION

Both sides have used credibility methods that are well-documented and widely used. How
can this discrepancy be resolved?

COMMENTARY

Both methods have arbitrary components. The regulator’s method required a prior
distribution. Had there been data available from a variety of insurers a different prior might
be more appropriate. The company had arbitrary choices when setting the standard for full
credibility. However, any set of reasonable choices will lead to assigning a small amount of
credibility to the company experience.

The Bayesian method assigns considerable credibility to the company experience not because
it views the data as particularly reliable, but because it considers the prima facie rate to also
be unreliable. This is the major difference in the two methods: the Bayesian method
requires evaluation of the credibility of the default value.

This problem also exposes a weakness of the limited fluctuation method when the true
probability is small. Recall that full credibility is expressed in terms of relative, not absolute
error. A 5% error means we get full credibility only if the estimate is within 0.00008 of the
true value. This degree of precision is far more than necessary.

An alternative is to consider the usual binomial confidence interval upper limit:

0.0016 + 1.645 0.004(0.996) / 5000 = 0.00307

Credibility Practice Note, July 2008

Page 37 of 55
which leads to a conservative premium of $6.14, which may be an acceptable compromise.
The prior probability was used to provide additional conservativism.

One way to reconcile these points of view is until the regulator can gather appropriate data to
have confidence in the prior distribution, focus on the one piece of information we have,
which is that company’s data and that can be used with the confidence interval approach.

Finally, the objectives of the regulator are quite different from those of the company actuary.
These differences should be understood in order to assist the actuary with understanding the
most appropriate approach used to resolve the given credibility issue.

Credibility Practice Note, July 2008

Page 38 of 55
4. A Short History and Background on Credibility Theory and its Usage in Actuarial
Practice

This section draws heavily from the article that Tom Herzog wrote on credibility for the
January/February 2008 issue of the Contingencies publication of the Academy.

INTRODUCTION

Credibility helps determine the weight that the actuary believes should be applied to a
particular body of experience in order to estimate future values. It is an example of a
statistical estimate. Statistical estimates are obtained through the use of statistical formulas
or models which, in turn, are based on statistical approaches or paradigms. There are
currently two main approaches or paradigms to statistical inference: (a) the frequentist
approach (also called sampling theory or the classical paradigm); and (b) the Bayesian
approach. We summarize each of these briefly in the next two sections.

FREQUENTIST APPROACH

Under the Frequentist Approach, the probability of an event is its relative frequency.
Random variables, such as the aggregate claim amount in a period of observation, are
assumed to have probability distributions whose parameters are constants. Prior
information enters statistical model building only in the selection of a class of models. The
main tools used are confidence intervals, unbiased estimates and tests of hypotheses.

BAYESIAN APPROACH

Under the Bayesian approach, probability is treated as a rational measure of belief. Random
variables, such as the aggregate claim amount in a period of observation, are assumed to have
probability distributions whose parameters may also have probability distributions. Recent
information and prior opinions and/or information (i.e., information available before the
recent information) are used to construct the probability distributions of the statistical models
or parameters of interest. Thus, the Bayesian approach is based on personal or subjective
probabilities and involves the use of Bayes’ theorem. Bayes’ Theorem says that if A and B
are events and P[B] > 0 , then

P[B | A ] ⋅ P[A ]
P[A | B] = (4.1)
P[B]

This theorem/approach was developed by the Reverend Thomas Bayes. Besides Bayes’
theorem, the main tools of Bayesian statistical inference are predictive distributions, posterior
distributions, and (posterior) odds ratios.

Credibility Practice Note, July 2008

Page 39 of 55
Inverse probabilities and statistical inference

Bayes’ Theorem is important to actuaries because it enables them to perform statistical

inference by computing inverse probabilities. We use the term “inverse” since we infer
backwards from results (or effects) to causes. Some simple examples follow.

A typical probability problem might be stated as follows: I have a standard die with six sides
numbered from “one” through “six” and throw the die three times. What is the probability
that the result of each of three tosses of this die will all be a “six”?

Now, I might have a second (non-standard) die with three sides numbered “one” and three
sides numbered “six”. Again I can ask the same question: What is the probability that the
result of each of three tosses of this die will all be a “six”?

The idea behind inverse probabilities is to turn the question around. Here, we might observe
that the results of three throws of a die were all “sixes.” We then ask the question: What is
the probability that we threw the standard die (as opposed to the non-standard die), given
these results?

Predictive Distributions

In insurance work, we typically experience a number of claims or an aggregate amount of

losses in one or more prior observation periods. Given these results we want to answer the
following questions:

(1) How many claims will we experience during the next observation period?
(2) What will be the aggregate loss amount during the next observation period?

Using Bayes’ Theorem, we can construct an entire probability distribution for such future
claim frequencies or loss amounts. Probability distributions of this type are called predictive
distributions. Predictive distributions give the actuary much more information than would an
average or other summary statistic. A predictive distribution provides the actuary with much
more information than just the expected aggregate amount of losses in the next period. For
example, it provides the actuary with a complete profile of the tail of the probability
distribution of aggregate losses as one might be concerned with in a “value-at-risk” analysis.
Thus, predictive distributions can provide the actuary and her client an important tool with
which to make business decisions under uncertainty.

BASIC CREDIBILITY CONCEPTS

Under some approaches to credibility, a compromise estimator, C, is calculated as

C = ZR + (1 − Z)H (4.2)

where R is the mean of the current observations (for example the data), H is the prior mean
(for example, an estimate based on the actuary’s prior data and/or opinion), and Z is the

Credibility Practice Note, July 2008

Page 40 of 55
credibility factor, satisfying the condition 0 ≤ Z ≤ 1. The credibility factor, Z, is the weight
assigned to the (current) data; (1-Z) is the weight assigned to the prior data.

As an insurance example, a new premium rate, C, is derived as a weighted average of an old

premium rate, H, and a premium rate, R, whose calculation is based solely on observations
from a recent period. An alternative interpretation is to let C be the premium rate for a class
of business, to let R be the premium rate whose calculation is based solely on the recent
experience of that class, and to let H be the insurance rate whose computation takes into
account the recent experience of all classes combined.

The credibility factor, Z, is typically of the form

n
Z= (4.3)
n+k

where k > 0 and n is a measure of exposure size. Under such a formulation, the problem
becomes how to determine k.

Over the years there have been three major approaches to credibility: Limited Fluctuation,
Greatest Accuracy and Bayesian. The Limited Fluctuation Approach and the Greatest
Accuracy Approach fall in the frequentist category of approaches. The Greatest Accuracy
Approach is also called the Buhlmann Approach, after Hans Buhlmann.

Limited Fluctuation Approach to Credibility

The Limited Fluctuation Approach seeks to calculate a compromise estimate, C, of the form
found in

C = ZR + (1 − Z)H . (4.2)

The limited fluctuation approach is one of the oldest, going back at least as far as Mowbray’s
1914 paper and Perryman’s 1932 paper. More modern treatments, including complete
mathematical derivations, are found in Longley-Cook’s 1962 article and Chapter 5 of
Herzog’s 1999 textbook. Outside of North America this approach is sometimes referred to as
“American Credibility.”

The main advantage of the limited fluctuation approach is that it is relatively simple to use.
However, a number of researchers/practitioners have raised questions about the limited
fluctuation approach to credibility.

Hans Buhlmann, as reported in Hickman and Heacock [1999], felt that the mathematical
reasoning behind the limited fluctuation approach as presented by Longley-Cook was “not
convincing.” Buhlmann noted that the approach:

1. Was under the frequentist paradigm of statistics and so ignored prior data.
2. Began by deriving the minimum volume of risks required for full credibility.

Credibility Practice Note, July 2008

Page 41 of 55
3. Relied on a derivation that was performed using a confidence interval.
4. Used a standard deviation argument to go from the full credibility requirement to partial
credibility for a smaller volume of risks.

Buhlmann raised the following concern with Longley-Cook’s derivation: Why should a
confidence interval that, by definition, includes the true value with a probability of less than
one, guarantee full credibility? Others have asked why the same weight, 1-Z, is given to the
prior mean, H, regardless of the analyst’s view of the accuracy of H.

Also, in the case of full credibility, no weight is given to the prior mean, H, as all of the
weight is given to the observed data, R. This raises the philosophical question of whether it
makes sense to talk about full (i.e., 100%) credibility because more data can generally be
obtained. Some analysts believe that no data are entitled to full credibility, so that the
credibility curve should approach 1 asymptotically, without ever reaching it.

Finally, according to Sundt [1991], there is an internal inconsistency in the limited

fluctuation approach to full credibility. He notes that the criterion for replacing the old
premium rate is based on the assumption that the old premium rate is correct. This leads to
the following conundrum: if the old premium is correct, then why replace it?

Buhlmann Approach (Greatest Accuracy Credibility)

The credibility estimates produced by Buhlmann’s approach are the best linear
approximations to the corresponding Bayesian point estimates. Under this approach, the
estimated parameters of the linear model are those that minimize the sum of squared
differences between the linear model and the observations. Because Buhlmann’s approach
typically

1. requires less mathematical skill,

2. requires fewer computational resources, and
3. does not require the selection of a prior distribution, which can sometimes require
difficult judgments,

Buhlmann’s approach to producing point estimates is often much more computationally

tractable than a complete Bayesian approach.

Bayesian Approach

Whitney [1918] stated that the credibility factor, Z, needed to be of the form

n
Z= (4.3)
n+k

where n represents “earned premiums” and k is a constant to be determined. The problem

was how to determine k. Whitney noted that, “In practice k must be determined by
judgment.” Specifically, the actuary needs to select the prior probabilities – i.e., the prior

Credibility Practice Note, July 2008

Page 42 of 55
 distribution. Whitney also noted that, “The detailed solution to this problem depends upon
the use of inverse probabilities” via Bayes’ Theorem.

Thoughts on the Choice of a Prior Distribution

Buhlmann expressed strong support for the Bayesian paradigm, but felt that whenever
possible the prior should be based on experience data rather than on subjective judgment. To
quote Buhlmann, as reported in Hickman and Heacox,

“Whereas early Bayesian statisticians used the prior distribution of risk parameters as a
means to express judgment (which in insurance we would call underwriting judgment), [I]
think of the probability distribution of the risk parameters as having an objective meaning. It
hence needs to be extracted from the data gathered about the collective. (Only in the case of
a lack of such data might one accept the subjective view faute de mieux.) For this reason, I
have always insisted on speaking about the structural distribution of risk parameters,
avoiding the standard Bayesian terminology, ‘prior distribution’.”

Fortunately, as Enrique de Alba points out (in his discussion of a paper by Udi Makov
[2001]) “actuarial science is a field where very frequently one has considerable prior
information, be it in the form of global or industry-wide information (experience) or in the
form of tables.” Because there is so much “objective” prior information available to the
actuary, de Alba is surprised that practicing actuaries have not made more extensive use of
Bayesian methods.

Bayesian Approach and Modern Computing

With the increased power of 21st-century computing equipment, advances in statistical

algorithms (e.g., the EM algorithm and Markov chain Monte Carlo methods) to implement
the Bayesian approach, and widely-available software that performs Bayesian inference (i.e.,
WinBUGS 5 ), a wider class of problems is becoming susceptible to solution via the Bayesian
approach.

Another Approach

Panel data, also known as longitudinal data, models are regression-type models that have
been applied widely in the biological and social sciences. Frees, Young, and Luo [1999]
show that many credibility models can be considered special cases of the longitudinal data
model. As a consequence, widely-available statistical software that has been produced to
analyze longitudinal data can be used to construct credibility models as well. Frees, Young,

5
The BUGS (Bayesian inference Using Gibbs Sampling) Project (begun by the MRC
Biostatistics unit at Imperial College, London) is concerned with the development of flexible
software for Bayesian analysis of complex statistical models using Markov chain Monte Carlo
methods. The “Win” prefix refers to Microsoft’s windows. For details see David Skollnik’s
2001 paper “Actuarial Modeling with MCMC and BUGS”.

Credibility Practice Note, July 2008

Page 43 of 55
and Luo [2001] have a number of examples that illustrate such applications to actuarial
problems.

Credibility Practice Note, July 2008

Page 44 of 55
5. Appendices

5.1 Bibliography of Recommended Reading on the Subject of Credibility Theory and Related
Approaches

5.2 Regulatory Guidance

Credibility Practice Note, July 2008

Page 45 of 55
5.1. Bibliography of Recommended Reading on the Subject of Credibility Theory and
Related Approaches

• Actuarial Standards Board (American Academy of Actuaries), Actuarial Standard

of Practice No. 25, “Credibility Procedures Applicable to Accident and Health,
Group Term Life, and Property/Casualty Coverages”, October 1996.

• Buhlmann, H. and Straub, E., “Credibility for Loss Ratios” (Translated by C.E.
Brooks), ARCH, 1972.2 (1972).

• Buhlmann, H. and Gisler, A., “A Course in Credibility Theory and its

Applications”, Berlin: Springer, 2005.

• Canadian Institute of Actuaries, Educational Note, “Expected Mortality: Fully

Underwritten Canadian Individual Life Insurance Policies”, July, 2002.

• Dean, C.G., “Topics in Credibility Theory”, Education and Examination

Committee of the Society of Actuaries, Construction and Evaluation of Actuarial
Models Study Note, 2005.

• Frees, Edward E., Young, Virginia R. and Luo, Yu, “A Longitudinal Data
Analysis Interpretation of Credibility Models,” Insurance: Mathematics and
Economics, 24 (1999), pages 229-247.

• Frees, Edward E., Young, Virginia R. and Luo, Yu, “Case Studies Using Panel
Data Models”, North American Actuarial Journal, Vol. 5, No. 4, October 2001.

• Herzog, T., “Introduction to Credibility Theory” (Third Edition), Winsted,

ACTEX Publications, (1999).

• Herzog, T., “The Rev. Thomas Bayes and Credibility Theory”, Contingencies
Magazine from the American Academy of Actuaries, January/February 2008.

• Hickman, J.C. and L. Heacox, “Credibility Theory: The Cornerstone of Actuarial

Science”, North American Actuarial Journal, Vol. 3, No. 2, April 1999.

• Klugman, S, “Sample size selection for multiple samples – A brief introduction to

credibility theory and an example featuring race-based insurance premium”, for
presentation at New York University, March 4, 2004.

• Klugman, Panjer and Willmot, “Loss Models: From Data to Decision”, Wiley-
Interscience, (2004).

Credibility Practice Note, July 2008

Page 46 of 55
• Longley-Cook, L.H., “An Introduction to Credibility Theory”, PCAS, 49 (1962),
pages 194-221.

• Mahler, H., “An Actuarial Note on Credibility Parameters”. Society of Actuaries,

Proceedings, Vol. LXXIII, Part 1, No. 139 (May 1986).

• Mayerson, A.L., “A Bayesian View of Credibility”, PCAS, 51 (1964), 85-104;

PCAS, 52 (1965), pages 85-103.

• Mowbray, A.H., “How Extensive a Payroll Is Necessary to Give a Dependable

Pure Premium?” PCAS, 1 (1914), pages 24030.

• Norberg, R. (London School of Economics), “Credibility Theory”, Encyclopedia

of Actuarial Science (2004), Wiley.

• Panjer, H and Hardy, M.R., “A Credibility Approach to Mortality Risk”, ASTIN

Bulletin, 1998, Vol. 28, No. 2, pages 269-283.

• Perryman, F.S., “Some Notes on Credibility”, PCAS, 19 (1932), pages 65-84.

• Sundt, B., “On Greatest Accuracy Credibility with Limited Fluctuation”,

Scandinavian Actuarial Journal, 2 (1992), pages 109-119.

• Venter, G., “Credibility Theory for Dummies”, Casualty Actuarial Society, CAS
Forum, Winter 2003, pages 621-627.

• Venter, G, (Mahler and Dean), “Credibility” (4th Edition), Chapter 8 of

Foundations of Casualty Actuarial Science. New York: Casualty Actuarial
Society, 1989.

• Whitney, A., “The Theory of Experience Rating”, PCAS, 5 (1918), 274-292.

Most of these books/papers are available at www.actexmadriver.com , on the
Casualty Society website at www.casact.org or the Society of Actuaries website at
www.soa.org (the last 2 websites provide the material free of charge).

Credibility Practice Note, July 2008

Page 47 of 55
5.2. Regulatory Guidance

5.2.1. Regulatory Guidance #1 from Florida (FAC 690-191.055, related to HMOs) -

excerpts

(1) (b)1. Pricing assumptions shall reflect assumptions based on sound actuarial
principles reflecting actual anticipated experience. Pricing assumptions shall be based
on the HMO experience to the degree credible.

(3) (b) 8. c. In determining medical trend, the HMO shall use credible data and make
appropriate adjustments to claims data to isolate the effects of medical trend only. This
shall not include the effects of underwriting wear off, aging, changes to claim costs due
to changes in demographics, contract coverages, geographic distribution, or
reinsurance.

(3) (b) 8. d. An HMO without fully credible data may, at its option, use an annual
medical trend assumption not to exceed the values in subsection XYZ for the medical
trend assumption without providing explicit trend justification.

(4) ( c) Credible Data :

1. If a contract form has 2,000 or more subscribers in force, then full (100%) credibility
is given to the experience; if fewer than 500 subscribers are in force, then zero (0%)
credibility is given. Linear interpolation is used for in force amounts between 500 and
2,000.

2. For group contract forms, the numbers in this definition refer to group subscribers.

3. Medical trend shall be used for the non-credible portion of the analysis.

Credibility Practice Note, July 2008

Page 48 of 55
5.2.2. Regulatory Guidance #2 from Colorado (3 CCR 702-4, Section 6, related to Rate
Filings and Annual Reports for Health Insurance) - excerpts

Section 6. Credibility: The State standard for fully credible data is 2,000 life years and
2,000 claims a year. Both standards must be met within a maximum of three years, if
the proposed rates are based on claims experience.

1. The memorandum must discuss the credibility of the State data with the proposed
rates based upon as much State data as possible. The memorandum must also
identify and discuss the source, applicability and use of collateral data used to
support partially credible State data. The use of collateral data is only acceptable if
the State data does not meet the full credibility standard. The formula for
determining the amount of credibility to assign to the data is SQRT{(#life years or
claims)/full credibility standard}. The full credibility standard is defined above.

2. The memorandum should also discuss how and if the aggregated data meets the
State credibility requirement. Any filing, which bases its conclusions on partially
credible data, should include a discussion as to how the rating methodology was
modified for the partially credible data.

Credibility Practice Note, July 2008

Page 49 of 55
5.2.3. Regulatory Guidance #3 from North Carolina (11 NCAC 16.0401) – Credit Life
Accident and Health Rate Deviation – excerpts

(3) " Case " means either a "Single Account Case" or a "Multiple Account Case" as
follows:

(a) " Single Account Case " means an account that is at least 25% credible or, at the
option of the insurer, any higher percentage as determined by the Credibility Formula
as defined in Item (6) of this Rule; and

(b) " Multiple Account Case " means two or more accounts of the same plan of
insurance and class of business having similar underwriting characteristics, excluding
single account cases defined in Sub-item (3)(a) of this Rule, and which, when
combined, are at least as credible as the minimum level of credibility elected in Sub-
item (3)(a) of this Rule.

(5) " Credibility Factor " means the degree to which the past experience of a case can
be expected to occur in the future.

(6) " Credibility Formula " means the following process used to calculate the credibility
factor:

(a) Determine the incurred claim count during the experience period;

(b) Divide Sub-item (6)(a) of this Rule by 1082;

(c) Take the square root of Sub-item (6)(b) of this Rule; and

(d) The credibility factor is the lesser of the number one and the results of Sub-item
(6)(c) of this Rule.

Credibility Practice Note, July 2008

Page 50 of 55
5.2.4. Regulatory Guidance #4 from Texas (28 TAC s 3.3307) – Minimum Standards for
Medicare Supplement Policies – excerpts

(d) (2) (E) (i) This rate change and demonstration shall be based on the experience of
the named form in this State only, if that experience is fully credible as set out in
paragraph (3) of this subsection.

(ii) The rate change and demonstration shall be based on experience of the named form
nationwide, with credibility factors as set out in paragraph (3) of this subsection
applied, if the named form is used nationwide and the experience in this State is not
fully credible.

(iii) The rate change and demonstration shall be based on experience of the named form
in this State only, with credibility factors as set out in paragraph (3) of this subsection
applied, if the named form is used in this State only and this State’s experience is not
fully credible.

(3) For purposes of this subsection, if a group or individual policy form has 2,000 or
more policies in force, then full credibility (100%) shall be given to the experience. If
fewer than 500 policies are in force, then no credibility (0%) shall be given to the
experience. The principle of linear interpolation shall be used for in-force numbers
between 500 and 2,000. For group policy forms, the reference in this paragraph to the
number of in-force policies means the number of in-force certificates under group
policies. For purposes of this section, " in force " means either the average number of
policies in force for the experience period used to support the need for a rate revision,
or the number of policies in force as of the ending date of the experience period used to
support the need for a rate revision. Once an issuer makes a decision as to which
definition it will apply to a particular policy form, such decision is irrevocable. An
issuer may submit specific alternate credibility standards to the department for
consideration. In order for an alternate standard of credibility to be acceptable for
application, the issuer must demonstrate that the standards are based on sound actuarial
principles, and that the resulting loss ratios are in substantial compliance with the
requirements of subsections (a),(b) and (c) of this section.

Credibility Practice Note, July 2008

Page 51 of 55
5.2.5. Regulatory Guidance #5 from California (Title 10, Chapter 5, subchapter 2,
Article 6.8) – excerpts

TEST FOR DEVIATED RATES

(a) An insurer shall use life insurance rates and disability insurance rates lower than the
prima facie rates for an experience group if the credibility adjusted loss ratio for that group
is equal to or less than the Presumptive Loss Ratio (PLR) less .05.
(b) An insurer may use life insurance rates and disability insurance rates higher than the
prima facie rates for an experience group when the credibility adjusted loss ratio for that
group equals or exceeds PLR + .05.
(c) For application of this Section to life insurance and disability insurance, credibility
adjusted loss ratios shall be computed according to Section XYZ for the most recent
experience period, based upon applicable prima facie rates for the experience group.

CALCULATION OF THE NEW CASE RATE

(2) “Experience period” means the most recent period of time for which experience is
reported, but not for a period longer than three full years. An experience period shall end
on December 31st of each calendar year.
(A) If an experience group develops a Credibility Factor of 1 from TABLE 4 (“Rate
Deviation Credibility Table” in Section XYZ) in less than three years, the experience
period for that case will be the number of full years needed to attain that Credibility
Factor.
(B) An insurer may elect a level of credibility within the range of TABLE 4 for an
experience group. If an experience group develops the minimum credibility elected by the
insurer in less than three years, the experience period for that group may be, at the option
of the insurer, either the number of full years needed to develop the elected minimum
credibility or three full years. If an experience group fails to attain such minimum
credibility within three full years, the credibility actually attained in that period shall be
used for determining downward deviated rates. Experience incurred in the period
immediately preceding the effective date of this regulation may be used to the extent
necessary to fill out the experience period.
(4) “Average Number of Life Years”

(c) For life insurance and for disability insurance, calculate the credibility adjusted loss
ratio(“CLR”) for the experience group using the following formula, where ALR is the
actual loss ratio for the experience group on the prima facie rate basis.
CLR = Z(ALR) + PLR(1-Z)
Calculate the new case rate (“NCR”) for the experience group according to the formulas
below, where PFR is the prima facie rate for the experience group.
(1) Where CLR is equal to or less than PLR less .05, the downward deviated rate may not
exceed the new case rate(“NCR”) calculated as:

Credibility Practice Note, July 2008

Page 52 of 55
 NCR = PFR [1 - (PLR-CLR)]
(2) Where CLR exceeds PLR + .05, the upward deviated rate may not exceed the new case
rate (“NCR”) calculated as:
NCR = PFR [1 + 1.2(CLR-PLR)]

TABLES
TABLE 4
RATE DEVIATION CREDIBILITY TABLE

Credibility
Average Number of Life Years Factor
(“Z”)
Life Disability Incurred
(all plans) Claim Count

14 Day 30 Day
1 1 1 1 .00
1800 141 209 9 .25
2400 188 279 12 .30
3000 234 349 15 .35
4600 359 535 23 .45
5600 438 651 28 .50
6600 516 767 33 .55
7600 594 884 38 .60
9600 750 1116 48 .65
11600 906 1349 58 .70
14600 1141 1698 73 .75
17600 1375 2047 88 .80
20600 1609 2395 103 .85
25600 2000 2977 128 .90
30600 2391 3558 153 .95
40000 3125 4651 200 1.00

The above integers represent the lower end of the bracket for each “Z” factor. The upper
end is 1 less than the lower end for the next higher “Z” factor.

Credibility Practice Note, July 2008

Page 53 of 55
5.2.6. Regulatory Guidance #6 from Maine (rule 220) – excerpts on Credit Life and
Health Insurance for one state

(1) The total deviated rate for a specified plan of benefits shall be the appropriate
promulgated prima facie premium rate increased or decreased by the additional
rate produced by the following formula:

Credibility Factor x (Actual/Expected Ratio - 1) x Prima Facie Claim Cost

(2) A case meets minimum credibility standards if the credibility factor from the
table below is at least equal to the minimum credibility factor elected by the
insurer. An insurer may make this election by notice to the Superintendent, in
writing. The minimum credibility factor elected may not be less than 50 percent.
Once elected, the minimum credibility factor will remain in effect for the insurer
until a different factor has been filed by the insurer and approved by the
Superintendent. If an insurer makes no written election, its minimum credibility
factor will be 100 percent.

(3) The credibility factor may be based on either the Number of Claims incurred or
on the Number of Life Years for the case during the experience period. The
insurer shall notify the Superintendent in writing which method it will use to
measure the credibility of all its cases in this State and may not change its
method without prior approval. If Claim Count or Life Year data is not available,
reasonable methods of approximation approved by the Superintendent may be
used until such data is developed.

Credibility Practice Note, July 2008

Page 54 of 55
 CREDIBILITY TABLE

Number of Life Years Incurred Credibility

Credit Life Credit Accident & Health Claim Count Factor
1- 1,799 1- 208 1- 8 .00
1,800 - 2,399 209 - 278 9- 11 .25
2,400 - 2,999 279 - 348 12 - 14 .30
3,000 - 3,599 349 - 418 15 - 17 .35
3,600 - 4,599 419 - 534 18 - 22 .40
4,600 - 5,599 535 - 651 23 - 27 .45
5,600 - 6,599 651 - 766 28 - 32 .50
6,600 - 7,599 767 - 883 33 - 37 .55
7,600 - 9,599 884 - 1,115 38 - 47 .60
9,600 - 11,599 1,116 - 1,348 48 - 57 .65
11,600 - 14,599 1,349 - 1,697 58 - 72 .70
14,600 - 17,599 1,698 - 2,046 73 - 87 .75
17,600 - 20,599 2,047 - 2,394 88 - 102 .80
20,600 - 25,599 2,395 - 2,976 103 - 127 .85
25,600 - 30,599 2,977 - 3,557 128 - 152 .90
30,600 - 39,999 3,558 - 4,650 153 - 199 .95
40,000 + 4,651 + 200 + 1.00

Credibility Practice Note, July 2008

Page 55 of 55