Marshall - Framework For Assessing Risk Margins

A Framework for Assessing Risk Margins
A Framework for Assessing Risk

Margins
Prepared by the Risk Margins Taskforce

(Karl Marshall, Scott Collings, Matt Hodson &
Conor O’Dowd)
This is the final version of a draft paper that was presented to the Institute of
Actuaries of Australia 16th General Insurance Seminar 9-12 November 2008, Coolum,
Australia. The changes between this and the draft are minimal and reflect our view
that the fundamental principles and techniques discussed in the draft remain
appropriate.
1
Contents
1. Introduction ...................................................................... 4
1.1. Preamble .................................................................. 4
1.2. Current approaches to assessing risk margins ............................................ 4
1.3. Practical framework for assessing risk margins .......................................... 6
1.4. Structure of this paper ............................................................................... 7
2. The proposed framework ............................................................................... 9

2.1. Introduction to framework ......................................................................... 9
2.2. Sources of uncertainty ............................................................................. 11
2.3. Preparing the claims portfolio for analysis ............................................... 12
2.4. Analysing independent risk sources .......................................................... 13
2.5. Analysing systemic risk sources ............................................................... 14
2.6. Consolidation of analysis into risk margin calculation ............................. 20
2.7. Additional analysis ........................................................................ 23
2.8. Documentation and regularity ................................................................... 27
3. Independent risk assessment ............................................................................ 29
4. Systemic risk assessment ............................................................................. 31

4.1. Internal systemic risk ............................................................................. 31
4.2. External systemic risk ............................................................................. 36
2
Abstract
The main purpose of this paper is to propose a comprehensive framework for assessing
insurance liability risk margins and to provide practical advice on how to implement it. The
key sources of uncertainty are examined and the main quantitative approaches to analysing
uncertainty discussed, including commentary on the advantages and disadvantages of each
approach. The framework recognises, however, that quantitative analysis of historical data
cannot alone capture adequately all aspects of future uncertainty. There will always be a need
for judgement to be applied and in many situations such considerations will dominate the risk
margin assessment. The application of judgement, however, is arguably the most difficult
aspect of any attempt to estimate future uncertainty and assess appropriate risk margins. Our
paper examines the key judgmental aspects and introduces a structured approach to
combining these qualitative considerations with the results of any available quantitative
analysis.
Keywords: framework, risk margins, uncertainty, APRA, independent risk, systemic risk.
3
1. Introduction
1.1. Preamble
General Insurance actuaries in Australia have, for many years, been analysing the
uncertainty involved in the claim process with a view to assessing appropriate risk
margins for inclusion in insurance liabilities. The approaches adopted to date range
from those that involve little analysis of the underlying claim portfolio to those that
involve significant analysis of the uncertainty using a wide range of information and
techniques, including stochastic modelling.
The Risk Margins Taskforce was created to provide GI actuaries in Australia with
support and guidance in the assessment of risk margins. In particular, it was felt that
actuaries would benefit greatly from a stronger awareness of the key considerations
when analysing uncertainty and the tools at their disposal when undertaking such
analysis. A better equipped actuarial profession could feel more confident that key
stakeholders, including APRA, insurance company boards, senior management and
auditors, better understand the nature of and feel more comfortable with the quality
and consistency of actuarial advice in this area.
The main purpose of this paper is to propose a comprehensive framework for

assessing insurance liability risk margins and to provide practical advice on how to
implement it. The key sources of uncertainty are examined and a combination of
quantitative and qualitative approaches to their measurement explored.
1.2. Current approaches to assessing risk margins
In preparation for a presentation to the 2006 Reserving Seminar of the Institute of

Actuaries of Australia (IAAust), the Taskforce canvassed a number of actuaries and
APRA to gain a better understanding of the range of approaches used in Australia to
assess risk margins. This information was supplemented with feedback from the
2006 General Insurance Claims Reserving and Risk Margins Survey, the results of
which were presented at the same seminar.
Although there appear to be a wide range of approaches used by Australian actuaries

in the assessment of risk margins it is fair to say that most of the differences relate to
the analysis and investigations conducted to parameterise a generally adopted risk
margin calculation methodology, rather than the calculation methodology itself. The
calculation methodology can be generalised as follows:
Coefficients of variation (CoVs) are determined for individual valuation

portfolios or groupings of portfolios, where these groupings include insurance
classes made up of relatively homogeneous risks.
A correlation matrix is populated with assumed correlation coefficients reflecting
the expected correlations between valuation portfolios or groupings of portfolios.
CoVs and correlation matrices are determined separately for outstanding claim
liabilities and premium liabilities and further assumptions made about the
correlation between these two components of the insurance liabilities.
A statistical distribution is selected and combined with the adopted CoVs and
correlation coefficients to determine the aggregate risk margin at a particular
probability of adequacy.
4
The approaches used to determine CoVs vary significantly. The least sophisticated
approaches involve deriving CoVs using either or both of two papers, Research and
Data Analysis Relevant to the Development of Standards and Guidelines on Liability
Valuation for General Insurance by Bateup and Reed (the Tillinghast paper) and
APRA Risk Margin Analysis by Collings and White (the Trowbridge paper), both
prepared at the end of 2001 (collectively these papers are referred to as the 2001
papers). These approaches often ignore the individual characteristics of the valuation
portfolio for which risk margins are being assessed, deferring instead to the
characteristics of the portfolios analysed by the authors of the two papers.
More sophisticated approaches include some form of quantitative analysis (stochastic

or otherwise) supplemented by a qualitative assessment of the sources of uncertainty
not captured by quantitative techniques. One such approach is discussed in the paper,
A Framework for Estimating Uncertainty in Insurance Claims Cost by O’Dowd,
Smith and Hardy, prepared for the IAAust’s XVth General Insurance Seminar which
was held in October 2005 (the PwC paper).
Anyone who has read the PwC paper will appreciate the similarities between the
framework proposed in that paper to the framework discussed in this paper. The
Taskforce is collectively of the view that the PwC paper has significant merit and
the concepts advocated by the authors of that paper have played a prominent
role in the development of the framework discussed in this paper. We would
encourage readers of this paper to read the PwC paper to ensure a more
complete understanding of some of the concepts discussed.
The most common approach to populating the correlation matrix with correlation
coefficients is via the deployment of actuarial judgement. Usually the key risks that
are considered to cause valuation portfolios to be correlated are considered in turn
and the correlation between classes categorised as high, medium or low with each
category having associated correlation coefficient values. The techniques deployed in
the assessment of correlations range from those that are quite basic and heavily
influenced by the benchmark correlation matrices discussed in the 2001 papers to
those that take a more methodical approach to analysing the contribution to
correlation from each key risk.
It is more the exception than the norm to include a quantitative analysis of past
experience in the assessment of correlation effects. The main reason for this is that
most quantitative techniques require a significant amount of data, time and cost to
produce results that are sufficiently credible and intuitively justifiable. It is more
common to see such techniques deployed when assessing more extreme probabilities
of adequacy, i.e. well in excess of 90%, rather than probabilities of adequacy around
the 75% level.
Generally, the most common distribution adopted to determine the aggregate risk
margin at a particular probability of adequacy is the LogNormal distribution. The
Normal distribution is also used by some actuaries, particularly at lower probabilities
of adequacy where it can generate a risk margin that is higher than a heavier tailed
distribution, such as the LogNormal distribution. It is uncommon for actuaries to test
the adopted distribution against past experience or, taking a step further, derive a
distributional form that explains the shape of the distribution of future claim cost
outcomes based on past experience and/or future expectations.
The general risk margins approach adopted by most actuaries is often referred to as a
bolt-on approach in that separate analyses are conducted to estimate the central
5
estimate of insurance liabilities and the risk margins. The term bolt-on is also
generally used to refer to any approach that does not involve the development of a
single unified distribution of the entire distribution of possible future claim cost
outcomes.
Judgement pervades both the central estimate assessment process and the risk margin
assessment process. Also, well fitting models are those that adequately reflect past
sources of uncertainty only. For these reasons, it is impossible to develop a purely
quantitative model, fitted to the past data, that accurately represents the range of
possible future claim cost outcomes. Rather, an approach that advocates internal
consistency between the assessment of the central estimate and the sources of future
uncertainty around that central estimate is important. The framework discussed in
this paper is one such approach. This transparent framework combines quantitative
and qualitative analysis, both of which are conducted giving full consideration to the
central estimate assessment.
1.3. Practical framework for assessing risk margins
A number of key stakeholders, including Appointed Actuaries, APRA and auditors,

have expressed some concern that the wide range of approaches adopted in practice to
assess risk margins might lead to significant inconsistencies in the final outcomes,
whether those be for regulatory or financial reporting purposes. Actuaries working in
this area have also asked for guidance to help them when they are faced with
analysing uncertainty. Finally, APRA have indicated that they would like to see more
documentary justification of the risk margins adopted by some insurance companies.
With all of this in mind, we have prepared this paper to provide a comprehensive
framework for assessing insurance liability risk margins and to provide practical
advice on how to use this framework. There are a number of parts to our framework
including the provision of guidance and further information on the tools, both
quantitative and qualitative, that an actuary may deploy when analysing the
uncertainty associated with insurance liabilities. We have included or referred to
practical examples of how to deploy parts of the framework.
The proposed framework recognises that quantitative analysis of historical data

cannot alone capture adequately all possible sources of future uncertainty. There will
always be a need for judgement to be applied and in many situations such
considerations will dominate the risk margin assessment. The application of
judgement, however, is arguably the most difficult aspect of any attempt to estimate
future uncertainty and assess appropriate risk margins. Our paper examines the key
judgmental aspects and introduces a structured approach to combining these
qualitative considerations with the results of any available quantitative analysis.
In preparing this paper the Taskforce has mainly considered, as a surrounding

context, the current risk margin environment in Australia, in particular the percentile,
or quantile, approach to determining margins for uncertainty. Having said this, we
are aware that international developments, including proposed changes to
International Financial Reporting Standards, are likely to overtake us in the not too
distant future. We are of the view that the main aspects of our proposed framework
can be readily adopted, altered or enhanced to complement analysis of uncertainty in
the evolving wider international context.
6
The framework discussed in this paper can also be considered in the broader context
of quantifying the uncertainty associated with reserve risk and underwriting risk for
stochastic capital modelling (often referred to as Dynamic Financial Analysis or
Internal Capital Modelling) purposes. In fact, when parameterising these
components of a DFA model, one should draw on any analysis conducted for risk
margin purposes and expand the framework to encapsulate those aspects of the
parameterisation not captured by an analysis conducted specifically for risk margin
purposes.
It is not proposed that this risk margin framework will have the prescriptive nature of
a professional standard. Nevertheless, it is hoped that the structure and educational
benefits it provides will encourage all actuaries to critically examine their current risk
margin methodologies and to take from the framework those insights that are helpful
to them in their particular situation. Inevitably, each actuary estimating risk margins
will need to make their own judgements and this will be driven by their own
knowledge and experience. The proposed framework does not attempt to usurp that
process. Ultimately this framework is about enabling the profession and stakeholders
to feel more confident in the quality and overall consistency of risk margins advice in
future.
This is not a paper on stochastic reserving. Nor is it intended to provide all of the
answers. Rather, its aim is to equip actuaries to ask the right questions and then
proceed to answer these in a methodical and rigorous manner.
1.4. Structure of this paper
In Section 2, we present a framework which takes a methodical and rigorous

approach to examining each of the key sources of uncertainty and provides a practical
and user-friendly platform to help actuaries determine appropriate and justifiable risk
margins for their insurance liability valuation portfolios.
Sections 3 and 4 discuss the assessment of independent risk and systemic risk,
respectively, providing more practical guidance and considerations for the assessment
of these sources of risk with a view to determining risk margins.
The framework is summarised in Table 1. The sections of the paper that address each
step are also shown.
7
Table 1: Summary of risk margin analysis framework

Step Framework component Description Section of paper
Determine valuation portfolios, claim groups and techniques to deploy

1 Portfolio preparation Section 2.3
for each claim group
Conduct quantitative analysis, conduct benchmarking where

2 Independent risk analysis Sections 2.4 and 3
appropriate, conduct retrospective analysis for stable periods
Apply balanced scorecard approach to objectively score central

3 Internal systemic risk analysis estimate valuation methodologies. Conduct analysis to determine Sections 2.5 and 4
appropriate CoVs to map to scores.
Identify, categorise and quantify potential future external sources of

4 External systemic risk analysis Sections 2.5 and 4
systemic risk
Select correlation coefficients beween valuation classes and between

5 Analysis of correlation effects outstanding claim and premium liabilities for internal systemic risk and Sections 2.5
for each external systemic risk category.
Consolidate CoVs and correlation coefficients. Independence assumed

6 Consolidation of analysis Section 2.6
between three sources of uncertainty.
Conduct sensitivity testing, scenario testing, internal and external

7 Additional analysis Section 2.7
benchmarking and hindsight analysis.
Document the analysis and judgement relating to each step of the

8 Documentation Section 2.8
framework
Conduct annual reviews of key assumptions in the context of emerging

9 Review experience. Full deployment of the framework at least every three Section 2.8
years, including active interactions with business unit management.
8
2. The proposed framework
2.1. Introduction to framework
The proposed framework provides a practical and robust platform that requires a
combination of quantitative and qualitative techniques to be deployed to examine the
uncertainty associated with assessing insurance liabilities with a view to determining
risk margins.
Quantitative techniques alone are insufficient to enable a complete assessment of the

various sources of uncertainty. These techniques must be supplemented by
qualitative analysis to ensure that all sources of uncertainty are captured. It is
common practice for Australian actuaries to adjust the results obtained using
quantitative techniques to allow for their known weaknesses. However, this is not
always done in a rigorous manner, nor is there much consistency across the
profession.
The framework is designed to introduce more rigour and consistency to the risk
margin assessment process by encouraging actuaries to examine their own portfolios
using a step-by-step process that requires them to ask a number of questions in the
context of these portfolios. This will enable judgemental aspects of the process to be
better reasoned, justified and documented and ultimately provide more structure in
the application and combination of both quantitative and qualitative processes.
It is not expected that all of the techniques discussed in this paper will be used in
practice for all valuation portfolios. Rather, if an actuary proceeds through the step-
by-step process using techniques suited to their own portfolios, understanding the
strengths and weaknesses of these techniques and asking the right questions along the
way, they can only be more comfortable that the risk margins adopted are
appropriate.
The framework revolves around quantifying the contribution to uncertainty from each
of the main sources of uncertainty and is graphically represented in Figure 1 below.
9
Figure 1: Framework for determining insurance liability risk margins

Claim s portfolio is the aggregate
XYZ Insurance
Claims Portfolio claims portfolio (e.g. licenced
Ltd.
insurance entity) for w hich the risk
margins must be estimated.
Valuation Classes represent the

portfolios that w ill be be considered
Valuation Class Valuation Class individually as part of the risk Home Motor
margin analysis. These may be aligned
to the valuation portfolios analysed
separately for central estimate purposes.
Claim Group is a group of claims

Homogeneous Homogeneous Attritional
homogeneous in terms of risk Liability claims Event claims
Claim Group Claim Group claims
characterisitics
System ic Risk is defined as risks

Independent Independent
Systemic Risk w hich are potentially common or Systemic Risk
Risk Risk
shared across Claim Groups or
These risks, w hen aggregated, Valuation Classes
make up the random com ponent
of param eter and process
External to Internal to uncertainty To ensure adequate identification External to Internal to
Actuarial Actuarial of causes of risk, Systemic Risk is Actuarial Actuarial
Process Process separated into risks external to Process Process
the actuarial process (external systemic
risk ) and risks internal to the
actuarial process (internal systemic
Risk Risk risk ) Labour Costs Specification
Risk Error
Parameter
Risk Risk Material Costs selection
Risk error
Risk Risk Frequency Data Error
These risks, w hen aggregated, make These risks, w hen aggregated, make
up the system ic com ponent up the system ic com ponent
of process uncertainty of param eter and m odel
uncertainty
10
2.2. Sources of uncertainty
The sources of uncertainty are the cornerstones of the framework. The framework
itself has been designed to ensure alignment between the analysis and the techniques
deployed with the key sources of uncertainty, ensuring a complete measurement of
uncertainty.
At the highest level, the sources of uncertainty can be categorised as belonging to

either the systemic risk source or the independent risk source.
Systemic risk represents those risks that are potentially common across valuation
classes or claim groups. Systemic risks arise from two sources:
Risks internal to the insurance liability valuation process, collectively referred to

in this paper as internal systemic risk. This source of uncertainty encapsulates the
extent to which the adopted actuarial valuation approach is an imperfect
representation of a complex real life process. Model structure and adequacy,
model parameterisation and data accuracy are all aspects of internal systemic risk.
This source of uncertainty is alternatively referred to as model specification risk.
Risks external to the actuarial modelling process, collectively referred to in this
paper as external systemic risk. Even if the valuation model is an appropriate
representation of reality, as it exists today, future systemic trends in claim cost
outcomes that are external to the modelling process may result in actual
experience differing from that expected based on the current environment and
trends.
Independent risk represents those risks arising due to the randomness inherent in the
insurance process. Independent risk also arises from two sources:
The random component of parameter risk, representing the extent to which the
randomness associated with the insurance process compromises the ability to
select appropriate parameters in the valuation models.
The random component of process risk being the pure effect of the randomness
associated with the insurance process. Even if the valuation model was perfectly
calibrated to reflect expected future outcomes, the volatility associated with the
insurance process is likely to result in differences from the perfect expected
outcomes.
In the detailed discussion of the framework below, quantitative and/or qualitative

techniques are considered and aligned to the assessment and measurement of the
internal and external sources of systemic risk and independent risk, the latter
incorporating both parameter and process risk.
The nature of traditional quantitative modelling techniques, e.g. bootstrapping and

stochastic chain ladder, are such that they are best suited to analysing sources of
independent risk and past episodes of external systemic risk. However, they are
inadequate alone to capture internal systemic risk or external systemic risk, to the
extent that this latter differs from the past. For both systemic risk sources, traditional
quantitative modelling techniques must be supplemented by other analysis, both
quantitative and qualitative.
11
2.3. Preparing the claims portfolio for analysis
Before commencing any analysis one must prepare the claims portfolio for analysis.
The claims portfolio would normally represent the aggregate insurance entity or
aggregation of insurance entities for which the risk margin analysis is being
conducted.
The claims portfolio should be split into appropriate valuation classes. A number of
factors will impact how the valuation classes are selected.
An important consideration is whether the valuation portfolio split adopted to

determine central estimates of insurance liabilities, or outstanding claim liabilities and
premium liabilities where the split is different, should be adopted for risk margin
analysis purposes. This would be preferable as it allows the risk margin analysis to
be conducted in the context of the central estimate analysis and quantitative and
qualitative analysis to be aligned with the key valuation drivers observed as part of
the central estimate valuation. One of the attractions of the framework is that each of
the sources of uncertainty being analysed can be aligned with the central estimate
analysis and appropriate decisions around volatility made in the context of that
analysis.
It may not be possible or particularly insightful, however, to conduct quantitative

analysis at the same granular level as used for central estimate valuation purposes.
The central estimate valuation portfolios may be too small for credible analysis or the
valuation portfolio allocation may be at a more granular level than makes practical
sense. For example, a large insurer may split its motor and home portfolios by state,
product and claim type, resulting in a large number of individual central estimate
valuation portfolios. The task of conducting quantitative analysis at the same
granular level may be significant, costly and, considering the level of qualitative
analysis that will be deployed as part of the assessment, unlikely to materially
improve the final outcome. In such cases, quantitative analysis may be conducted on
aggregated valuation classes and the results then allocated down, in an appropriate
manner, to the valuation classes that are considered appropriate for the deployment of
the framework.
In the end, the choice of valuation classes for risk margins analysis purposes will
come down to a balance between the practical benefits gained from a higher level
portfolio allocation and the potential additional benefit and insights gained from a
more granular allocation. When making this decision consideration should be given
to the need to retain as much consistency as possible between the central estimate
methodology and basis and the risk margin analysis.
Once the claims portfolio has been allocated into risk margin valuation classes,
consideration should be given to whether any valuation classes would benefit from a
further allocation. For certain portfolios, it will be apparent that different groups of
claims are materially more or less uncertain than others and should be treated
separately for risk margin analysis purposes. Within each of these claim groups there
is an element of homogeneity but between claim groups behaviour is expected to be
different.
A good example of a valuation class that would normally require further segregation
is a home portfolio. These portfolios are normally materially exposed to claims
arising from natural peril events. The patterns of development for event claims often
differ materially from those for non-event claims. Separate analysis of event and
12
non-event claims will usually provide valuable insights into the past contribution to
uncertainty from each of these claim sources with a view to making appropriate
assumptions regarding future uncertainty. Also, home liability claims typically
behave quite differently from other home claims and should be considered for
separate analysis.
Again, a pragmatic view should be taken when considering whether groups of claims
are homogeneous, a view that balances the benefits against the practicalities and cost.
For certain valuation portfolios, e.g. those with little historical data, it may not be
possible to deploy all components of the framework. However, we do consider it
important to consider each component in the context of each valuation portfolio as
this will ensure that appropriate questions are asked as part of the analysis.
2.4. Analysing independent risk sources
Many approaches used in practice by actuaries to analyse uncertainty and assess risk
margins have an element of quantitative analysis conducted using stochastic (or
other) modelling techniques. Often, but not always, adjustments are made to the
results from this modelling, reflecting an appreciation that it has not fully
encapsulated all sources of uncertainty.
There are a number of reasons why stochastic modelling techniques do not enable a
complete analysis of all sources of uncertainty:
A good stochastic model will fit the past data well and, in doing so, fit away most
past systemic episodes of risk external to the valuation process, leaving behind
largely random sources of uncertainty. Some techniques, e.g Generalised Linear
Modelling (GLM), offer more flexibility in fitting to the past experience than
others, e.g. Mack method.
Where it has not been possible to fit away all past systemic episodes of risk or
where no attempt has been made to do so, the outcome of the analysis may be
substantially affected by these episodes. Consideration then needs to be given to
whether past episodes of systemic risk are reasonably representative of what one
can expect in the future. For some portfolios this will be a very significant
assumption, based solely on judgemental considerations.
Even where one is comfortable that a model adequately reflects the volatility
expected in the future from both independent and systemic sources external to the
actuarial valuation process, the model is highly unlikely to incorporate
uncertainty arising from sources internal to the actuarial valuation process, i.e.
internal systemic risk.
The framework proposes the use of one or more stochastic modelling techniques to
analyse independent sources of risk and to inform on past episodes of systemic risk
external to the actuarial valuation process. There are a number of approaches that
may be used to analyse independent sources of risk, including:
Mack method;
Bootstrapping;
Stochastic Chain Ladder;
Generalised Linear Modelling (GLM) techniques; and
Bayesian techniques.
13
Although these techniques can be used for both outstanding claim liabilities and
premium liabilities, it is possible and practically helpful to analyse independent risk
as it pertains to premium liabilities using techniques specifically designed for this
purpose.
The analysis of independent risk is an art in itself and actuaries will only become
comfortable in this area with practical experience of working through the main issues
on their own valuation portfolios. A range of stochastic techniques may be used and
decisions made on the strengths and weaknesses of each approach in the context of
the past experience. It may be possible to refine the modelling to focus on certain
past periods with limited past episodes of systemic risk, thus largely isolating past
independent risk and examining the extent to which it has impacted past volatility.
Finally, we do consider it useful to supplement any analysis of independent risk for a

particular valuation portfolio with internal and external benchmarking.
Benchmarking is discussed in section 2.7. The main source of external benchmarking
in this regard would be the 2001 Tillinghast paper which identified the independent
risk component in its overall uncertainty benchmarks. For some portfolios,
benchmarking may be the only way to obtain some view of the contribution from
independent risk once all other avenues have been exhausted.
2.5. Analysing systemic risk sources
The framework proposes separate analysis of internal systemic risk and external
systemic risk. Qualitative approaches are proposed for this purpose. Two approaches
are discussed in Section 4 of the paper, one designed to analyse internal systemic risk
and the other designed to analyse external systemic risk. Introductions to these
approaches are given in this sub-section. Both techniques have been designed to
allow judgement to be deployed in a robust, transparent and consistent manner,
giving due consideration to each of the key contributors to the two sources of
systemic risk.
Internal systemic risk
Internal systemic risk refers to the uncertainty arising from the actuarial valuation
models used being an imperfect representation of the insurance process as it pertains
to insurance liabilities. Valuation models are designed to predict future claim cost
outcomes based largely on an examination of the key predictors of claim cost, and
trends in these predictors, as these have been observed in the past claim experience.
When assessing the uncertainty associated with the insurance liabilities it is important
to subject the valuation methodology to objective scrutiny to assess the extent to
which the quality of the insurance liability estimate may be compromised by
inadequacies in the valuation process. The need to be objective as part of this process
is important. Human nature is such that it is easy to become overly defensive of the
modelling approach adopted for central estimate purposes. Objective comparisons
and scoring of the adopted valuation methodology against best practice, irrespective
of whether such best practice is possible in the context of the portfolio being
analysed, is crucial to forming an appropriate view of the contribution of internal
systemic risk to uncertainty.
We consider there to be three main sources of internal systemic risk. These are:
14
Specification error - the error that can arise from an inability to build a model
that is fully representative of the underlying insurance process. The process is
likely to be too complicated to be replicated in any actuarial valuation model.
Also, the information available may be such that the underlying process cannot
be fully understood and the model structure is simplified as a consequence.
Parameter selection error - the error that can arise because the model is unable to
adequately measure all predictors of claim cost outcomes or trends in these
predictors. Again the insurance process is such that there can be a large number
of claim cost drivers that would be difficult to fully capture in an actuarial
valuation model.
Data error - the error that can arise due to poor data or unavailability of data
required to conduct a credible valuation. Data error also relates to inadequate
knowledge of the portfolio being analysed, including pricing, underwriting and
claims management processes and strategies.
One approach to analysing internal systemic risk is discussed in detail in section 4 of

the paper. This involves developing a balanced scorecard to objectively assess the
model specification against a set of criteria designed to rank aspects of the modelling
from worst to best practice. For each of the sources of internal systemic risk, risk
indicators are developed and then scored against the adopted criteria. The scores are
then aggregated for each valuation class and mapped to a quantitative measure (CoV)
of the variation arising from internal systemic risk.
There are a number of subjective decisions that are required to be made as part of this
process. These include the risk indicators, the measurement and scoring criteria, the
importance (or weight) afforded to each risk indicator and the CoVs that map to each
score from the balanced scorecard. Quantitative techniques may be used to inform
aspects of these decisions.
Development and deployment of a balanced scorecard approach to measuring internal

systemic risk is a blend of art and science. Actuaries unfamiliar with the approach
will need time to develop the skills required:
to draw out all of the risk indicators;

objectively score them against best practice; and
map them to a CoV in the context of their own valuation classes.
Section 4 of the paper provides some thoughts and tools that may be used as part of
such an exercise. However, it is fully expected that new techniques will emerge as
experience develops and the writers of this paper welcome and encourage future
contributions to the development of actuarial thinking in this area.
The analysis of internal systemic risk is summarised in Figure 2 below.
15
Figure 2: Internal systemic risk – systemic risk internal to the actuarial valuation process
Source of Risk
Qualitative assessment
uncertainty component Combine scores Map scores to CoV
Assess specification error using a range of risk

indicators and diagnostics
Risk indicators may include number of models

Specification used and range of results, reasonableness checks
error conducted, subjective adjustments required,
extent of monitoring and review.
Score each risk indicator against best practice

using a range from 1 to 5
Assess parameter selection error using a range of Weighted average valuation class
risk indicators and diagnostics scores mapped to CoVs.
A weighted average combined score Low scores will attract high CoVs
Risk indicators may include ability to identify and for each valuation class is derived. high scores low CoVs.
Parameter use predictors, extent to which predictors lead Weights subjectively selected to CoVs may differ between long-tail
Internal
selection rather than lag claim costs, subjective adjustments refect the actuary's view of the and short-tail portfolios and between
systemic risk
error required, ability to detect trends, stability, importance of each of the risk outstanding claims and premium
uncertainty in superimposed inflation. indicators in the context of their own liabilities.
portfolio. CoVs derived based on a
Score each risk indicator against best practice combination of judgement and
using a range from 1 to 5 analysis.
Assess data error using a range of risk indicators

and diagnostics
Risk indicators may include extent, timeliness and

Data error reliability of information from business, access to
data, quality of reconciliations, extent of revisions
to past data.
Score each risk indicator against best practice

using a range from 1 to 5
16
External Systemic Risk
All of the standard quantitative modelling techniques analyse the volatility inherent in
the past claim experience. As such, they can only be used to inform on the
uncertainty arising from past episodes of external systemic risk. To use these
techniques in isolation would require an assumption that the contribution to volatility
from future external systemic risk is expected to be similar to that experienced in the
past. It is quite possible, and for some valuation classes likely, that future external
systemic risk will exhibit significantly different characteristics from actual past
episodes.
It is, therefore, important to identify each of the main potential sources of external
systemic risk and, for each of these sources, quantify their impact on the overall
volatility of the insurance liabilities. The main external systemic risks for any
valuation class can be categorised as belonging to a number of risk categories. These
include:
Economic and social risks – normal inflation and other social and environmental
trends
Legislative, political risks and claim inflation risks – relates to known or
unknown changes to legislative or political environment within which each
valuation portfolio currently operates and shifts or trends in the level of claim
settlements (this risk category encapsulates most systemic trends normally
referred to as superimposed inflation)
Claim management process change risk – changes to the processes relating to
claim reporting, payment, finalisation or estimation
Expense risk – the uncertainty associated with the cost of managing the run off of
the insurance liabilities or the cost of maintaining the unexpired risk until the date
of loss
Event risk – the uncertainty associated with claim costs arising from events,
either natural peril events or man-made events
Latent claim risk – the uncertainty associated with claims that may arise from a
particular source, a source that is currently not considered to be covered
Recovery risk – the uncertainty associated with recoveries, either reinsurance or
non-reinsurance
Each of these risk categories will normally have been considered as part of the central
estimate valuation of outstanding claim or premium liabilities. There is, therefore, a
strong case for conducting the analysis of external systemic risk in conjunction with
the central estimate valuation, thereby ensuring that both parts of the valuation take a
consistent and complete view of all systemic risk categories.
A critical step in any valuation process is the interaction between the valuation
actuary and business unit management. This is required to ensure that the valuation
actuary has an appropriate level of understanding of all aspects of the insurance
process, particularly as this relates to the valuation of insurance liabilities. These
interactions will normally incorporate discussions about all aspects of the portfolio
management process, including underwriting and risk selection, pricing, claims
management, expense management, emerging portfolio trends and the environment
within which the portfolio operates. It would be of great benefit to the valuation
process, and not particularly onerous, to extend discussions to consider the main
potential external systemic risks that may impact the portfolio. This information can
17
then be used to inform both the central estimate valuation and in the identification
and quantification of risks associated with each external systemic risk category.
For most valuation classes, the risk identification and categorisation process will
identify a small number of systemic risks and categories that account for the majority
of the uncertainty. For property classes, for example, event risk is likely to dominate
the volatility of the premium liabilities whereas for long-tail portfolios legislative,
political and claims inflation risks are likely to be the key contributors to the
volatility for both outstanding claim and premium liabilities.
When analysing external systemic risk it is useful to rank each of the risk categories
in descending order in terms of expected impact on insurance liability uncertainty.
This ranking can then be used to guide the effort to be expended on quantifying the
risks associated with each risk category. More time and effort would be spent on
quantifying the uncertainty associated with material risk categories.
Section 4 of the paper discusses the assessment of external systemic risk in more
detail and includes some examples of potential sources of systemic risk within each
risk category.
Correlation effects
At this point in the deployment of the framework, an actuary will have derived CoVs
for independent risk, internal systemic risk and for each source of external systemic
risk in each systemic risk category. The next step requires making allowance for the
fact that each of these sources of risk is not fully correlated either within valuation
classes or between valuation classes.
At this stage, it is worth commenting that we do not consider or discuss any

quantitative methods to assessing correlation effects as part of this paper. The main
reasons for this are as follows:
Available techniques tend to be technically complex and often require a

substantial amount of data. The time and effort required to learn, implement and
appropriately adjust these techniques may outweigh the benefits gained.
These techniques will yield correlations that are heavily influenced by the
correlations, if any, experienced in past data. Correlations associated with
external systemic risk sources may differ materially from correlations associated
with past episodes of systemic risk.
Also, it is difficult, if not impossible, to separate the past correlation effects
between independent risk and systemic risk or to identify the pure effect of each
past systemic risk.
Internal systemic risk cannot be modelled using standard correlation modelling
techniques.
Even if modelling of correlation effects were practical, they are unlikely to yield
results that could be aligned to the outcomes of the framework discussed above in
relation to independent risk, internal systemic risk and external systemic risk.
Having said this, it is not our intention to entirely rule out quantitative analysis of past
correlation effects. Such analysis may provide useful insights that can help in the
assessment of potential future correlation effects.
18
The framework can be readily extended to incorporate an appropriate allowance for

correlation effects. This extension follows the spirit of the framework discussed so
far and requires that correlation effects be considered in the context of each source of
uncertainty and/or risk category. Again, reliance is placed on an actuary’s own
judgement but the actuary is encouraged to deploy their judgement in a robust and
transparent manner in the context of each of the risks affecting their valuation classes.
Correlation effects can be considered in the context of each source of uncertainty.

The key considerations are discussed below.
Independent risk – as suggested by the name, this source of uncertainty can be

assumed to be uncorrelated with any other source of uncertainty, either within a
particular valuation class or between valuation classes.
Internal systemic risk – this source of uncertainty can be assumed to be
uncorrelated with independent risk, as discussed above, and with each potential
external systemic source of risk, either within a particular valuation class or
between valuation classes. Internal systemic risk contributes to correlation
effects through correlation of this source of uncertainty between valuation classes
or between outstanding claim and premium liabilities.
o The same actuary effect and the use of template or valuation models across
different valuation classes are key considerations for correlation effects
between valuation classes.
o Linkages between the premium liability methodology and outcomes from the
outstanding claim valuation are key considerations for correlation effects
between outstanding claim and premium liabilities.
External systemic risk – it is reasonable to assume that the contribution to
uncertainty from each risk category is uncorrelated with independent risk,
internal systemic risk and with the contribution to uncertainty from each other
risk category, either within a particular valuation class or between valuation
classes. Correlation effects will arise from correlations between classes or
between outstanding claim and premium liabilities from risks categorised as
belonging to similar risk categories, e.g. claims inflation risk across long-tail
portfolios or event risk across property and motor portfolios.
It is possible that external systemic risk categories may be partially correlated either
within or between valuation classes. If this is the case, the correlated risk categories
may be aggregated into broader categories that are not correlated with other risk
categories.
For practical purpose, the correlation relationship between any two sources of
uncertainty or risk categories can be considered to belong to one of a finite number of
assumed correlation bands. For example, five correlation bands may be defined as
nil, low, medium, high and full correlation. For quantification purposes one might
allocate correlation coefficients of 25%, 50% and 75%, respectively, to the low,
medium and high correlation bands. Having any more than five categories is likely to
result in spurious accuracy attaching to what is already a largely subjective process.
The PwC paper describes a useful way of considering and assessing correlation
effects. A root dummy variable, which can be considered to be the root source of
correlations within a risk category, is created. Dummy variables may also be set up
for groupings of valuation classes that belong to the same class of business, e.g.
separate valuations may be conducted by state within a worker’s compensation class
19
of business. A hierarchical structure can then be constructed for each systemic risk
category containing correlations between the following components:
premium liabilities and outstanding claim liabilities for a particular valuation

class;
outstanding claim liabilities for individual valuation classes and the relevant class
of business dummy variables; and
class of business dummy variables and root dummy variables.
The implied correlations, both within valuation classes or classes of business and
between valuation classes, can then be assessed.
2.6. Consolidation of analysis into risk margin calculation
Once an actuary has progressed through the analysis discussed above they will have
the following assumptions that need to be consolidated and converted into a risk
margin for the whole claims portfolio:
CoVs in respect of independent risk for each valuation portfolio, separately for
outstanding claim and premium liabilities
CoVs in respect of internal systemic risk for each valuation portfolio, separately
for outstanding claim and premium liabilities
CoVs in respect of each potential external systemic risk category, separately for
outstanding claim and premium liabilities
Correlation coefficients between each source of uncertainty, risk category,
valuation portfolio and outstanding claim/premium liability combination.
For practical purposes, we propose that a simple linear correlation dependency

structure be adopted to allow for the various correlation effects. Correlation matrices
are created for each of the three sources of uncertainty described in section 2.2 above.
As discussed above, independent risk, internal systemic risk and external systemic
risk are all assumed to be uncorrelated. As such, the contribution from each source of
uncertainty to the total CoV, after correlation effects, can be calculated individually
and then combined.
We consider a simple linear correlation dependency structure to be reasonable for the

assessment of risk margins associated with probabilities of adequacy of up to at least
90%. Where one is faced with requirements for extreme probabilities of adequacy,
e.g. for portfolios in run off or when parameterising reserve risk for DFA modelling
purposes, it is recommended that other dependency structures be considered.
An example of the consolidation and risk margin calculation for an example insurer,
Insurer ABC, which underwrites three classes of business, Motor, Home and CTP is
shown in Figure 3 below.
20
Figure 3: Claims portfolio CoV and risk margin calculation for Insurer ABC
A: Proportion of insurance liabilities
Proportion of insurance liabilities

(weights)
Outstanding
Class claims Premium liabilities
Motor 5% 25%
Home 5% 25%
CTP 30% 10%
Total 40% 60%
B: Independent risk
Independent risk
Outstanding Insurance
Class claims CoV Premium liabilities CoV liabilities CoV
Motor 7.0% 5.0% 1.3%

Home 6.0% 5.0% 4.3%
CTP 6.0% 15.0% 5.9%
Total 4.6% 3.9% 3.0%
C: Internal systemic risk

Internal systemic risk Internal systemic risk correlation matrix
Motor OSC Motor PL Home OSC Home PL CTP OSC CTP PL
Motor OSC 100% 75% 50% 50% 25% 25%
Motor 5.5% 5.0% 4.9% Motor PL 75% 100% 50% 50% 25% 25%
Home 5.5% 5.0% 4.9% Home OSC 50% 50% 100% 75% 25% 25%
CTP 9.5% 8.0% 8.7% Home PL 50% 50% 75% 100% 25% 25%
Total 7.6% 4.2% 4.9% CTP OSC 25% 25% 25% 25% 100% 75%
CTP PL 25% 25% 25% 25% 75% 100%
D: External systemic risk

Economic, External systemic risk - coefficients of variation by risk category
social, etc, Legislative, political Claim process Latent claim All risk
risk and claims inflation risk risk Expense risk Event risk risk Recovery risk categories
Motor OSC 1.0% 0.5% 2.0% 1.0% 1.0% 0.0% 3.0% 4.0%
Motor PL 2.0% 0.5% 2.0% 2.0% 3.0% 0.0% 5.0% 6.8%
Home OSC 1.0% 1.0% 2.0% 1.0% 2.0% 0.5% 0.5% 3.4%
Home PL 2.0% 1.0% 2.0% 2.0% 15.0% 0.5% 1.0% 15.5%
CTP OSC 3.0% 10.0% 4.0% 2.0% 0.0% 0.5% 1.0% 11.4%
CTP PL 4.0% 12.0% 4.0% 3.0% 1.0% 0.5% 2.0% 13.8%
External systemic risk - risk category correlations

Risk category Correlations adopted
Economic, social and environmental risk Nil between CTP and other, 25% PL/25% OSC between motor and home, 50% between OSC and PL within classes
Legislative, political and claims inflation risk Nil between CTP and other, 25% PL/25% OSC between motor and home, 50% between OSC and PL within classes
Claim management process risk 25% between classes, 50% between OSC and PL within classes
Expense risk 25% between classes, 50% between OSC and PL within classes
Event risk Nil between CTP and other, 50% PL/25% OSC between motor and home, 50% between OSC and PL within classes
Latent claim risk Nil between classes, 50% between OSC and PL within classes
Recovery risk Nil between classes, 50% between OSC and PL within classes
External systemic risk

Motor 4.0% 6.8% 6.0%

Home 3.4% 15.5% 13.1%
CTP 11.4% 13.8% 10.7%
Total 8.6% 8.0% 6.5%
E: Consolidated CoVs
All sources of uncertainty
Motor 9.8% 9.8% 7.9%

Home 8.8% 17.0% 14.6%
CTP 16.0% 21.9% 15.0%
Total 12.4% 9.9% 8.7%
F: Risk margins
Required probability of adequacy 75%
Risk margins - Normal distribution Risk margins - LogNormal distribution
Outstanding Insurance Outstanding Premium Insurance
Class claims Premium liabilities liabilities claims CoV liabilities CoV liabilities CoV
Motor 6.6% 6.6% 5.3% 6.3% 6.3% 5.1%

Home 5.9% 11.5% 9.9% 5.7% 10.5% 9.2%
CTP 10.8% 14.8% 10.1% 9.9% 13.0% 9.4%
Total 8.4% 6.7% 5.8% 7.9% 6.3% 5.6%
21
The following comments are made to help in the interpretation of the example in
Figure 3.
The CoVs and correlation coefficients used and risk margins derived are
indicative only. The emphasis is on demonstrating how consolidation could work
in practice, rather than proposing appropriate risk margins or underlying
assumptions.
Part A gives the percentage breakdown of the total net central estimate of
insurance liabilities by valuation portfolio and between outstanding claim and
premium liabilities. There is no need to use actual dollar amounts in the
calculation. The percentage breakdown (or weights) will suffice. For simplicity,
for this example all homogeneous claim groups have been combined within the
valuation classes.
Part B shows the CoVs adopted in respect of independent risk for outstanding
claim and premium liabilities following a combination of quantitative modelling
and benchmarking. The insurance liability CoVs by valuation portfolio and the
insurance liability, outstanding claim liability and premium liability CoVs for all
valuation portfolios combined have been derived assuming independence (or nil
correlation) between valuation portfolios and between outstanding claims and
premium liabilities.
Part C shows the CoVs and correlation coefficients (in correlation matrix form)
adopted for outstanding claim and premium liabilities in respect of internal
systemic risk. These CoVs and correlation coefficients have been derived
following a qualitative analysis of internal systemic risk using a balanced
scorecard approach. The insurance liability CoVs by valuation portfolio and the
insurance liability, outstanding claim liability and premium liability CoVs for all
valuation portfolios combined have been derived using the assumed correlations
between valuation portfolios and between outstanding claim and premium
liabilities. When creating any correlation matrix it is important to include a check
that the matrix is positive definite.
The first table in Part D shows the CoVs adopted in respect of each external
systemic risk category. The second table summarises the adopted correlation
coefficients in respect of external systemic risk. The implementation of these
correlations is conducted using seven correlation matrices, one for each external
systemic risk category. Each of these matrices is 6x6, similar to the correlation
matrix shown in Part C for internal systemic risk. With an assumption of
independence between risk categories there is no need to create a larger 42x42
matrix with a row and column representing each risk category, valuation portfolio
and outstanding claim/premium liability combination. The CoVs and correlation
coefficients shown in these two tables have been derived following a qualitative
analysis of potential external systemic sources of risk. The third table in Part D
shows the aggregate CoVs in respect of external systemic risk, derived for each
valuation portfolio and for all valuation portfolios combined in respect of
outstanding claim liabilities, premium liabilities and insurance liabilities.
Part E consolidates the CoVs from each of the three sources of uncertainty,
derived in Parts B to D. The key assumption underlying the derivation of
consolidated CoVs is that there is independence between each of the sources of
uncertainty.
Part F converts the consolidated CoVs into risk margins assuming a required
probability of adequacy of 75%. Two statistical distributions have been adopted
as representative of the underlying distribution of insurance liabilities: the
Normal distribution and the LogNormal distribution. At lower probabilities of
adequacy, including 75%, the Normal distribution delivers a higher risk margin,
22
irrespective of the consolidated CoV. At higher probabilities of adequacy,

including 90%, the LogNormal distribution can give a higher result, where the
consolidated CoV is not too high. For particularly high CoVs, the LogNormal
distribution can generate risk margins that appear unreasonable. For example, for
a 75% probability of adequacy the risk margin percentage does not increase much
above 25% and actually reduces as the CoV increases above 75%. Another way
of looking at this is that LogNormal risk margins can reduce quite significantly as
a percentage of the CoV as the latter increases whereas Normal risk margins
remain unchanged as a percentage of the CoV.
Both distributions are used in practice by actuaries with the LogNormal
distribution more common for higher probabilities of adequacy and the Normal
distribution, for the reasons discussed above, often given consideration at the
75% probability of adequacy. The right-tailed nature of the distribution of
insurance liabilities perhaps lends itself more to a right-skewed distribution such
as LogNormal. However, it does have its practical issues at lower probabilities of
adequacy as discussed above. Considering the level of judgement required in the
application of the framework, spending a substantial amount of time deliberating
over the form of the distribution is unlikely to be of much value. An actuary
should adopt a distribution that is appropriate in the context of their own claims
portfolio, including the consolidated CoV assessed and probability of adequacy
required. One might not be so comfortable to adopt a LogNormal or Normal
distribution without further justification if the purpose of the analysis is to derive
risk margins with very high probabilities of adequacy (i.e. 99.5% for portfolios in
run off) or when parameterising reserve risk in a DFA modelling context.
A spreadsheet tool has been created to do the calculation required for the
consolidation shown in Figure 3. This tool has been provided as an attachment to
this paper to help readers understand the key formulae underpinning the
consolidation. Obviously, this tool may also be adapted for use in the
deployment of the framework discussed in this paper.
2.7. Additional analysis
There are a number of areas of additional analysis that may be conducted to give an
actuary further comfort regarding the outcomes from the deployment of the
framework described above. These include sensitivity analysis, scenario testing,
benchmarking and hindsight analysis, each of which is discussed below.
Sensitivity testing
The framework requires a substantial amount of actuarial judgement in its

application. Judgement is required in all aspects of the analysis, irrespective of
whether quantitative or qualitative methods have been used to assess the volatility
associated with a particular source of uncertainty.
Valuable insights into the sensitivity of the final outcomes to key assumptions can be
gained by varying each of the key assumptions. It is recommended that, as part of the
analysis, the CoVs and correlation coefficients adopted for independent risk, internal
systemic risk and each external systemic risk category be flexed and the impact on
the valuation class and claims portfolio risk margins examined.
23
Following such an analysis, one might review certain key assumptions, particularly
those that have a substantial impact on the final outcome, with a view to gaining
additional comfort that the adopted assumptions are reasonable and justifiable.
As a demonstration of sensitivity testing in practice changes have been made to

certain key assumptions adopted for the example in Figure 3.
If the independent risk CoVs by valuation portfolio for outstanding claim and
premium liabilities are reduced by 50%, the risk margin for the whole claims
portfolio (based on the LogNormal distribution) reduces from 5.6% to 5.4%.
Alternatively, doubling these CoVs increases the risk margin to 6.5%.
If the internal systemic risk CoVs by valuation portfolio for outstanding claim
and premium liabilities are increased by 50%, the risk margin for the whole
claims portfolio increases from 5.6% to 6.6%. Alternatively, increasing the
correlation coefficients to give full correlation across all combinations increases
the risk margin to 6.3%.
If the CoVs for the legislative, political and claims inflation systemic risk
category for CTP (outstanding claims and premium liabilities) are reduced by
50%, the risk margin for the whole claims portfolio reduces from 5.6% to 5.2%.
Doubling the CoV for the event systemic risk category for Home premium
liabilities increases the risk margin to 7.0%. Finally, assuming full correlation,
within all valuation classes and systemic risk categories, between outstanding
claim and premium liabilities increases the risk margin to 5.8%.
Scenario testing
It is often insightful to tie the risk margin outcomes back to a set of valuation
outcomes by strengthening some of the key assumptions adopted for central estimate
purposes to align the outstanding claim liabilities and premium liabilities with the
provisions assessed including risk margins. Various different assumption scenarios
may be tested and valuation outcomes, including projected ultimate claim
frequencies, average claim sizes, loss ratios, etc, compared for each scenario against
the central estimate basis.
These (risk margin inclusive) valuation outcomes can be considered in the context of
the emerging experience and what is known about the portfolio. Also, the basis
changes required to deliver these outcomes can be considered in the context of the
emerging experience.
Internal benchmarking
As part of the CoV selection process, the proposed CoVs should be subjected to a
range of internal checks. For each source of uncertainty individually the adopted
CoVs should be compared between valuation classes, particularly similar valuation
classes, for outstanding claim liabilities, premium liabilities and insurance liabilities.
Comparisons should also be made between outstanding claim and premium liability
CoVs within classes.
For independent risk, there are two main dimensions that should be considered in the
context of internal benchmarking: portfolio size and length of claim run off. The law
of large numbers implies that the larger the portfolio, the lower the volatility arising
from random effects. Also, the longer a portfolio takes to run off, the more time there
24
is for random effects to have an impact. These considerations have a number of

implications for independent risk CoV selection, including:
Outstanding claim liability CoVs for short-tail portfolios are likely to be lower
than for similar sized long-tail portfolio and substantially lower than much
smaller long-tailed portfolios.
Premium liability CoVs for long-tail portfolios would normally be higher than
outstanding claim liability CoVs for the same portfolios. This is due more to the
law of large numbers than any material differences in the length of the run off.
The extent of the difference will depend on the size of the premium liability and
outstanding claim liability with the difference being more for small portfolios
which will have higher independent risk components than for large portfolios
which will have smaller independent risk components.
Premium liability CoVs for short-tail portfolios would normally be lower than
outstanding claim liability CoVs for the same portfolios, assuming the same
independent risk profile between outstanding claim and premium liabilities. This
is due mainly to the law of large numbers. The independent risk profiles may
not, however, be similar. Event risk, where material, is likely to mean that the
independent risk profile of premium liabilities and outstanding claim liabilities
are different. This is likely to offset the benefit that premium liabilities gain from
their greater size and in any event make benchmarking problematic.
For internal systemic risk, the CoVs can be compared in the context of each valuation
class. If template models are used for similar portfolios, particularly classes with
homogeneous claim groups, then one would expect CoVs to be similar between
classes. Also, the underlying process and the key drivers of this process are likely to
be more complicated in long-tail portfolios than most short-tail portfolios. If similar
valuation methodologies are applied for both short- and long-tail classes then one
would expect higher internal systemic risk CoVs for the long-tail portfolios.
The main sources of external systemic risk are likely to be much more significant for
long-tail portfolios with the exception of event risk for property and, to a lesser
extent, motor classes and liability risk for home classes.
External benchmarking
External benchmarking refers to the use of the Tillinghast and Trowbridge 2001
papers or APRA’s November 2008 General Insurance Risk Margins Industry Report
to benchmark CoVs and/or risk margins derived as part of a risk margins analysis.
APRA have indicated that a large number of actuaries rely, to varying extents, on the
analysis presented in the 2001 papers in the selection of their own risk margin
assumptions. This reliance ranges from those actuaries who conduct thorough
analyses on their own portfolios and then benchmark the adopted risk margins with
those derived from the 2001 papers to those actuaries that derive risk margins solely
from the 2001 papers with little or no consideration of the reasonableness of this
approach in the context of their own portfolios. The latter approach was certainly not
one of the original intentions of the authors of the 2001 papers. The former approach
is more consistent with the expectations of the authors.
It is not our intention to dismiss external benchmarking out of hand. Rather, we

consider that this form of benchmarking has some merit when combined with a
thorough analysis of a particular claims portfolio. Benchmarking will be of some
25
benefit where there is little information available for analysis purposes, particularly
for the analysis of independent risk. More generally, the use of benchmarking should
be as a sanity check rather than as the entire basis of the risk margin assessment. In
any deployment of benchmarking, the differences between the benchmark portfolio(s)
and the claims portfolio being analysed must be considered and factored into the
analysis.
The use of the Tillinghast paper in the assessment of independent risk is discussed in
section 2.4 above. Before using the Tillinghast paper, however, an actuary needs to
be aware of the following issues:
The assumptions required to derive the independent component of the CoV were
derived based on an analysis conducted during 2001. The independent CoVs
depend on the size of the outstanding claim or premium liabilities. Inflation
between 2001 and the effective date of the current valuation should be backed out
of the outstanding claim and premium liabilities before calculating the
independent CoV. If this is not done then the independent CoV will be
understated.
The premium liability risk margin should be calculated by applying a multiple to
the outstanding claim risk margin for an outstanding claim liability that is the
same size as the premium liability, not for the actual outstanding claim liability,
irrespective of whether this is lower or higher than the premium liability.
Hindsight analysis
Hindsight analysis involves comparing past estimates of outstanding claim liabilities

and premium liabilities against the latest view of the equivalent liabilities.
Movements can be analysed and converted to a coefficient of variation reflective of
the actual volatility observed in the past. This volatility contains a combination of
past instances of independent risk, internal systemic risk and external systemic risk.
Care needs to be taken in the interpretation of any hindsight analysis as the models
may have changed (improved) since previous valuations were conducted. Also,
future external sources of systemic risk may differ materially from past such episodes
of systemic risk.
Hindsight analysis is particularly useful for short-tail valuations where there is little
serial correlation between consecutive valuations. Hindsight analysis is somewhat
less valuable for long-tail portfolios where there is usually significant serial
correlation between consecutive valuations.
The reader is referred to the 2005 paper An Empirical Approach to Insurance

Liability Prediction Error With Application to APRA Risk Margin Determination by
Andrew Houltram for a thorough discussion of the benefits and practicalities
associated with hindsight estimation.
Another form of hindsight analysis, which we will refer to as mechanical hindsight

analysis, is one that takes a mechanical approach to estimating the outstanding claims
and premium liabilities, systematically removing the most recent claims experience.
An example of such an approach is as follows:
Apply a chain ladder method on a triangulation of cumulative claim payments

based on a triangulation of data at the valuation date.
26
The adopted payment development factors should be calculated using an

objective approach, e.g. the average of the actual experience over the last three
years.
The outstanding claim payments derived using all data to the valuation date is
referred to as the ‘current’ estimate.
Remove a diagonal of payment data one at a time and apply the same method
objectively to derive outstanding claim payments at past valuation dates.
Compare each of the past estimates of outstanding claim payments with the
current estimate, for the equivalent accident periods and ensuring that relevant
payments made between valuation dates are added to the current estimate of
outstanding claim payments.
The method can be extended to incorporate a mechanical projection of premium
liabilities at each valuation date. Premium liability volatility and past levels of
correlation between outstanding claim and premium liabilities can be examined.
Mechanical hindsight analysis may be used to analyse:
independent risk, by focusing the analysis on periods where there was a degree of
stability in the experience with few or no systemic trends;
internal systemic risk, by applying this technique using a range of actuarial
methods (preferably those used for central estimate valuation purpose) and
observing the differences in volatility outcomes; and
all past sources of uncertainty, by applying the approach across all past periods.
The latter is a mechanical variant of the hindsight analysis described in the first three
paragraphs of this sub-section.
2.8. Documentation and regularity
Documentation
APRA have indicated that a wide range of approaches have been taken by actuaries in
the documentation of risk margins analysis. Documentation ranges from that which
provides a thorough discussion of approach and justification for the assumptions
underpinning the adopted risk margins to that which provides very little commentary
or justification.
Documentation of actuarial judgement is not necessarily an easy task. However, we

believe that the framework offers actuaries an opportunity to document their analysis
and key judgemental decisions in a complete and robust manner, aligned to the key
steps in the framework.
Regularity and review
A full application of each step of the framework is a substantial and comprehensive

undertaking. We do not consider that the framework need be applied in its entirety
each time an actuary conducts a central estimate valuation of insurance liabilities.
We consider a full application of the framework at less regular intervals to be

reasonable and appropriate. At the very least, however, a full application should be
applied every three years. These extensive reviews should incorporate all of the steps
27
of the framework discussed above and summarised in Table 1. They will also involve
significant interaction with business unit management.
At more regular intervals, aligned to the times when central estimate valuations of
insurance liabilities are conducted, a less comprehensive review of the key
assumptions adopted as part of the previous full application will suffice. The key
assumptions should be examined in the context of:
any emerging trends;

emerging systemic risks; and
changes to valuation methodologies.
Changes to key assumptions would only be considered where there is reasonable

justification for doing so, i.e. where the previous assumptions are no longer deemed
appropriate. Another way of thinking of these regular reviews are as monitoring
exercises where key assumptions derived from the previous full framework
application are monitored against emerging experience and developing knowledge
and adjusted where justified.
If new portfolios emerge in the period between full applications of the framework,
one should consider applying the key steps within the framework to those portfolios.
The successful deployment of this framework will require significant interaction with
business unit management. The process may benefit from a feedback and
communication loop, enabling the business to provide their views on the outcomes of
the analysis. This will reduce the possibility that lots of assumptions, which all make
sense individually, contribute to an overall outcome that does not make sense. This
communication loop may incorporate the demonstration of scenarios that would give
rise to the outcome assessed at the selected probability of adequacy.
28
3. Independent risk assessment
Independent risk reflects the contribution to the uncertainty associated with the actual
claim cost outcome from random effects. This source of risk has two components:
the random component of parameter risk and the random component of process risk.
It is not normally particularly enlightening or beneficial to split independent risk
between these two components. Having said this, some quantitative modelling
techniques do allow the split to be assessed as part of their normal application.
There are a number of approaches that may be used to analyse independent sources of
risk, including::
Mack method;
Bootstrapping;
Stochastic Chain Ladder;
Generalised Linear Modelling (GLM) techniques; and
Bayesian techniques
The bibliography includes references to a number of papers that describe these

techniques.
The techniques vary in their capacity to enable actuaries to identify past levels of
independent risk. In the application of most of these techniques, one is attempting to
fit a model to past systemic episodes and trends and to analyse the residual volatility
once these episodes and trends have been fitted away. The better the model fit is the
more likely that the residual volatility observed reflects random effects alone.
An actuary faced with the task of assessing independent risk will need to decide upon
which techniques to use for each of their valuation classes. This decision should
consider the extent to which the independent risk for a particular valuation class is
material to the overall claim portfolio risk margin, the contribution to uncertainty
from internal systemic risk and external systemic risk and the cost and effort
associated with applying the techniques. Where the cost and effort outweighs the
potential benefit then a simpler approach, perhaps incorporating benchmarking, may
be considered.
For some valuation portfolios, the data available may be too limited or volatile to
enable a credible split between past episodes of systemic risk and past independent
risk. In these cases, actuaries may consider using a model that does not attempt to fit
away the past systemic risk and supplement this analysis with additional allowances
for external systemic risk, to the extent that this is considered to differ from past
systemic risk, and internal systemic risk, which cannot be modelled using standard
quantitative modelling techniques.
Independent risk assessment for outstanding claim liabilities
Any of the techniques mentioned above can be used in the assessment of past
independent risk for outstanding claim purposes. Some of the techniques offer more
flexibility in terms of fitting to past systemic episodes and trends.
Consideration should be given to aligning the methodology adopted to analyse

uncertainty with that used for central estimate purposes. For example, if the PPCI
method plays an important role in the central estimate assessment and bootstrapping
29
is the preferred approach to analysing uncertainty then the PPCI method should be
bootstrapped. This will ensure that past volatility is examined and conclusions drawn
in an environment that is internally consistent.
GLM techniques can be used to model individual claims or aggregate claims. These
techniques are used for reserving purposes to identify the key factors that have
contributed to past claim cost outcomes. Combined with a range of useful statistical
diagnostics, these techniques are well placed to support the analysis of independent
risk.
Bootstrapping techniques offer less flexibility than GLM techniques but can be
adapted to help in the assessment of random effects. For example, if past periods that
have been largely unaffected by systemic episodes can be identified then the
bootstrap residuals can be calculated for these stable periods and used as part of the
bootstrapping process. Plots of residuals by accident period, development period and
experience period can be used to identify periods that have been affected by past
systemic episodes.
Independent risk assessment for premium liabilities
The bootstrapping, GLM and Bayesian approaches may also be used for the purpose
of analysing the volatility in past claim experience for the purpose of assessing the
independent risk component for premium liabilities.
However, it is possible to use simpler techniques to analyse the past volatility of key
components of the premium liabilities. Consider a valuation class where the central
estimate of the claim cost component of the premium liabilities is assessed by
combining a projected claim frequency and average claim size. The adopted claim
frequency and average claim size has been selected following an analysis of output
from the outstanding claim valuation supplemented by portfolio level pricing
analysis.
For some valuation classes, it can be a relatively straightforward exercise to remove

the impact of past systemic episodes (including seasonality) from observed claim
frequencies and determine the claim frequency CoV in respect of past residual
volatility. Similarly, past average claim sizes can be adjusted to remove past
inflation, including both standard and superimposed, and other past systemic episodes
(again including seasonality) and a CoV in respect of past residual volatility derived.
Where a loss ratio approach to projecting premium liabilities is used, allowance

should be made for systemic shifts in past premium levels as well as claim costs.
Often large claims are extracted for separate analysis. Again, observations can be
made as to the aspects of past experience that represent systemic episodes and those
that are purely random.
The process of identifying and isolating past systemic episodes can only be enhanced
if an actuary has a strong understanding of the possible systemic sources of risk for a
particular portfolio. The role that product and claim management can play in
improving this understanding should not be underestimated. This is discussed further
in section 4.
30
4. Systemic risk assessment
4.1. Internal systemic risk
Internal systemic risk refers to the uncertainty arising from the actuarial valuation
models used being an imperfect representation of the insurance process as it pertains
to insurance liabilities.
As discussed in section 2.5, we consider there to be three main sources of internal

systemic risk. These are:
Specification error - the error that can arise from an inability to build a model
that is fully representative of the underlying insurance process.
Parameter selection error - the error that can arise because the model is unable to
measure all predictors of claim cost outcomes or trends in these predictors.
Data error - the error that can arise due to poor data, unavailability of data and/or
inadequate knowledge of the portfolio being analysed.
When an actuary conducts an assessment of outstanding claim or premium liabilities,

there are a wide range and variety of approaches and methodologies that are
available. The merits of each approach will be considered in the context of the
valuation classes being assessed. The characteristics of each class and the level of
information available, including granularity of data, will all play a role in the decision
around which approach to use.
Although care will normally be taken to ensure that the approach adopted is
appropriate for the valuation class being assessed, models are likely to represent a
simplified view of the insurance process. Models also range in their capacity to
identify underlying trends in the claims experience. Standard triangulations methods
will normally analyse predictors (e.g. claim payments, reports, finalisations, case
estimates) that have been aggregated to a reasonably high level or lag rather than lead
the underlying drivers of the insurance process.
In light of this, any analysis of uncertainty would be incomplete without an objective

assessment of the adequacy of the modelling infrastructure and its ability to reflect
and predict the underlying insurance process. In this section of our paper, we propose
one approach, involving the development of a balanced scorecard, which may be used
as part of such an assessment.
One other point worth making before we walk through the balanced scorecard
approach in detail is that the assessment of internal systemic risk must be conducted
in the context of the actual approach used to assess the central estimate of outstanding
claim and premium liabilities. The strengths and weakness associated with that
approach will be considered and scored with a view to determining an appropriate
allowance in risk margins for internal systemic risk. Consistency between the central
estimate and risk margin assessments are one outcome of a robust assessment of
internal systemic risk.
The balanced scorecard was discussed in section 2.5 and presented diagrammatically
in Figure 2. In summary the approach involves:
For each of the specification, parameter and data risk components, conduct a
qualitative assessment of the modelling infrastructure, considering a range of risk
31
indicators and scoring these indicators on a scale of 1 to 5 (where 5 represents

best practice).
Apply weights to each risk indicator, reflecting its relative importance to the
overall modelling infrastructure, and calculate a weighted average score
representing an objective view of the quality of the modelling infrastructure for
each valuation class.
Calibrate the weighted average score derived to a CoV in respect of internal
systemic risk. The development of appropriate CoVs will likely involve a
substantial amount of judgement, perhaps supplemented by quantitative analysis.
In a paper entitled Asbestos Liabilities & the New Risk Margins Framework, prepared
by Brett Riley and Bruce Watson, the authors describe an alternative approach to
assessing the level of internal systemic risk. This approach specifies High and Low
scenarios that ‘represent the end points of what might be considered a reasonable
range of central estimates based on alternative interpretations of all available
information’. The approach advocated by Messrs Riley and Watson certainly has
merit and represents a reasonable alternative to the balanced scorecard approach
described in this paper. It also has the appeal of being simpler and, therefore, more
practical to apply.
Scoring the modelling infrastructure
We would encourage actuaries to develop a balanced scorecard approach that is

suited to the characteristics of the valuation classes within their own claims portfolio
including risk indicators that are most relevant in the context of these classes. Having
said this, we feel that it is useful if we outline potential risk indicators that actuaries
may wish to consider and develop for the purpose of their own analysis. Table 2
includes potential risk indicators and some suggested minimum requirements for a
high score for each of these indicators. The characteristics that represent a poor score
should be readily apparent.
32
Table 2: Internal systemic risk - Potential risk indicators
Risk component Potential risk indicators Requirements for high score
Many different modelling approaches considered - each approach should add value by considering different dimensions of
Number of independent models used
claims experience
Extent to which models separately analyse different claim/payment types Relevant homogeneous claim or payment types modelled separately
Low variations between different models in terms of past performance - take care that comparisons are appropriate (e.g. PCE
Range of results produced by models
vs PPCI for old accident periods for short-tail classes may not be appropriate)
Significant reasonableness checks conducted, including reconciliation of movement in liabilities, diagnostic checks on
Checks made on reasonableness of results
valuation outcomes, acceptance of results by business, expert peer review, benchmarking against industry
Confidence in assessment of model 'goodness of fit' Actual vs Extected close, few difficulties in selecting parameters, relevant sensitivities yield small variances in results
Specification error
Few subjective adjustments, relevant subjective factor sensitivites yield low variances and adjustments regularly monitored
Number and importance of subjective adjustments to factors
and reviewed
Extent of monitoring and review of model and assumption performance Model and assumption performance monitored continuously and reviewed regularly
Ability to detect trends in key claim cost indicators Models have performed well in detecting trends in the past
Sophistication and performance of superimposed inflation analysis Detailed analysis of past sources of superimposed inflation and robust quantification of each past source
Level of expense analysis to support CHE assumptions Detailed expense analysis, including how expenses are incurred over the lifetime of claims relating to each claim type
Ability to model using more granular data, e.g. unit record data Unit record data is available and used to further analyse and better understand key predictors and trends in these predictors
Best predictors have been analysed and identified, including internal and external variables that show strong correlaton with
Best predictors have been identified, whether or not they are used
claims experience
Parameter
Best predictors are stable over time or change due to process changes Predictors stable over time, stabilise quickly and respond well to process changes
selection error
Predictors are close to best predictors, lead (rather than lag) claim cost outcomes, modelled rather than subjectively allowed
Value of predictors used
for and unimpaired by past systemic events
Knowledge of past processes affecting predictors Good and credible knowledge of past processes, including changes to processes
Regular, complete and pro-active two-way communication between valuation actuary and claims staff/portfolio managers who
Extent, timeliness, consistency and reliability of information from business
understand key valuation predictors and how changes may impact or invalidate these
Reconciliations against other sources are conducted for all data sources and types, checks are conducted throughout data
Data error Data subject of appropriate reconciliations and quality control
processing steps, reconciliations against previous valuation conducted, data and differences well understood
Processes for obtaining and processing data are robust and replicable No past instances of poor data understanding, no or low potential for miscoding of claim type
Frequency and severity of past mis-estimation due to revision of data No past instances of data revision
Extent of current data issues and possible impact on predictors No known current data issues
33
Each of the risk indicators should be considered in the context of both the outstanding
claim and premium liabilities. Additional indicators may be considered for premium
liabilities, for example whether the outstanding claim liabilities are used as an input
to the premium liability assessment or whether credible portfolio level pricing
analysis is used as an input to the premium liability assessment.
For certain short-tail portfolios, some risk indicators may not be as relevant for
premium liability purposes. A large variance in the outstanding claim liabilities,
which might only affect the most recent accident periods and have a relatively small
impact on the projected ultimate claim frequency or average claim size, may not be
material in the context of a premium liability assessment.
Table 3 shows the risk indicator scores which underpin the internal systemic risk
CoVs adopted for Insurer ABC in the example in Figure 3 in section 2.6, with a
particular focus on outstanding claim liabilities.
Table 3: Internal systemic risk – example balanced scorecard

Motor Motor Home Home CTP score CTP
Risk component Potential risk indicators score OSC weight score OSC weight OSC weight
Number of independent models used 4 7 4 7 3 2

Extent to which models separately analyse different claim/payment types 3 3 4.5 5 2 7
Range of results produced by models 4 5 4 4 2 2
Checks made on reasonableness of results 5 5 5 5 4 5
Confidence in assessment of model 'goodness of fit' 4 5 4 5 2 7
Specification error Number and importance of subjective adjustments to factors 5 3 4 3 3 5
Extent of monitoring and review of model and assumption performance 4 5 4 5 5 8
Ability to detect trends in key claim cost indicators 4 4 3 4 3 6
Sophistication and performance of superimposed inflation analysis 0 0 4 10
Level of expense analysis to support CHE assumptions 4 4 4 4 2 2
Ability to model using more granular data, e.g. unit record data 2 2 2 2 5 2
Best predictors have been identified, whether or not they are used 4 3 4 5 3 7
Parameter
Best predictors are stable over time or change due to process changes 5 5 4 5 2 6
selection error
Value of predictors used 4 5 4 5 3 5
Knowledge of past processes affecting predictors 4 8 4 8 4 8

Extent, timeliness, consistency and reliability of information from business 4 5 4 5 4 5
Data subject of appropriate reconciliations and quality control 4 7 4 7 4 8
Data error
Processes for obtaining and processing data are robust and replicable 5 3 5 3 5 3
Frequency and severity of past mis-estimation due to revision of data 5 3 3 3 5 5
Extent of current data issues and possible impact on predictors 4 3 5 3 5 3
Total weighted average score - outstanding claims (OSC) 4.1 4.0 3.5
Total weighted average score - premium liabilities 4.5 4.5 4.0
The scores and weights shown in Table 3 are for illustration only and should be taken
as a demonstration of concept than as a set of benchmarks that actuaries can use for
such portfolios in practice.
The weights allocated to each of the risk indicators are a measure of the importance
of that risk indicator, relative to the other risk indicators, in terms of its contribution
to overall internal systemic risk. The weights and hence relativities between risk
indicators should reflect the particular valuation infrastructure adopted for each
valuation class including the relative importance of each risk indicator in the context
of that valuation class.
Premium liabilities scored better than outstanding claims in this example due to the
extensive use in their assessment of outcomes from the valuation of outstanding
claims and independent and credible portfolio level pricing analyses conducted
recently.
34
Calibrating scores to CoVs
Once a score representing an objective and qualitative view of the efficacy of the
modelling infrastructure has been derived, one needs to determine a CoV that is an
appropriate representation of the contribution to outstanding claim and premium
liability uncertainty from internal systemic risk. This step is likely to require a
significant amount of subjective judgement, supplemented by quantitative analysis.
We suggest that individual actuaries develop a CoV scale which represents their view
of the uncertainty associated with internal systemic risk for the full range of possible
balanced scorecard outcomes, ranging from worst practice to best practice (or
‘perfect’) modelling approaches. A large degree of judgement will be required to
derive a reasonable range in the context of a particular claims portfolio. The analysis
conducted to score the modelling infrastructure together with past model performance
should provide invaluable insights into the potential variability associated with a
particular modelling approach.
If more than one methodology has been deployed in the past then a hindsight analysis
of the actual past performance of each method can be used to assess the relative
performance of each method and the extent to which multiple models can improve the
performance of the whole modelling infrastructure.
Mechanical hindsight analysis (see section 2.7) may also be used to help in the
assessment of internal systemic risk. For example, a mechanical hindsight analysis
can be conducted using one method with all claim or payment types aggregated. A
further retrospective analysis can be conducted using multiple methods with claim or
payment types separated into individual homogeneous groups. The relative difference
in performance of the two modelling infrastructures over time may give some insights
into the additional uncertainty associated with poor modelling approaches compared
to fair or good modelling approaches.
Based on our experience, we would suggest that the minimum CoV associated with a
‘perfect’ model is unlikely to be much less than 5%. Even a ‘perfect’ model will not
be able to completely replicate the true underlying insurance process or identify every
possible predictor of claim cost outcomes.
If you consider a single, aggregated model with limited data or information available
to populate the model, significant subjective assumptions required and few identified
predictors, CoVs of 20% or above in respect of internal systemic risk are readily
justifiable. For such models, it is not infeasible that internal systemic risk could be
the main contributor to overall uncertainty.
Table 4 gives CoV scales used in the assessment of risk margins for Insurer ABC as
part of the example in Figure 3.
35
Table 4: Internal systemic risk – example CoV scale

Score from
balanced scorecard
assessment Motor CoV Home CoV CTP CoV
1.0 to 1.5 17.5% 17.5% 25.0%

1.5 to 2.0 13.0% 13.0% 20.5%
2.0 to 2.5 10.5% 10.5% 17.0%
2.5 to 3.0 8.5% 8.5% 14.0%
3.0 to 3.5 7.0% 7.0% 11.5%
3.5 to 4.0 6.0% 6.0% 9.5%
4.0 to 4.5 5.5% 5.5% 8.0%
4.5 to 5.0 5.0% 5.0% 7.0%
The CoV scale shown in Table 4 is an example only. Actuaries should select CoV
scales that are appropriate in the context of their own valuation classes and the
modelling infrastructure adopted for each of those valuation classes. Any hindsight
analysis deployed to support the selection of appropriate CoVs should be designed to
align with the actual valuation methods adopted for the valuation classes being
analysed.
Further comments on the CoV scale as presented in Table 4 are:
The scale is not linear reflecting our view that the marginal improvement in
outcomes between fair and good modelling infrastructures is less than the
marginal improvement between poor and fair modelling infrastructures.
The CoVs for CTP, a long-tail portfolio, are higher than those for Motor and
Home, both short-tail portfolios. For long-tail portfolios, it is generally more
difficult to develop a modelling approach that is representative of the underlying
insurance process. Also, key predictors are often less stable for long-tail
portfolios and past episodes of systemic risk more likely to impair the ability to
fit a good model.
The scale has been used for both outstanding claim and premium liability
purposes. A reasonable ‘a priori’ assumption is that similar scales can be used
for both. Arguments can be made for premium liabilities to have higher or lower
CoVs than those applying to outstanding claim liabilities, particularly for poor
modelling approaches. For example, the assessment of premium liabilities may
include additional uncertainty associated with the estimation of exposure or
premium relating to unclosed or contractually bound future business. If this is the
case then a loading on top of the outstanding claim liability CoVs may be
justifiable. On the other hand, for certain stable short-tail classes, the difference
between a simple loss ratio approach and a more thorough frequency/severity
approach may not be material in terms of performance in the assessment of
premium liabilities but the difference between a single aggregate model and
multiple disaggregated models could be material in terms of performance in the
assessment of outstanding claim liabilities.
4.2. External systemic risk
External systemic risk refers to the uncertainty arising from non-random risks
external to the actuarial modelling process. This uncertainty encapsulates systemic
episodes that have not yet occurred but may emerge in the future and those that are
emerging in the recent experience but where there is some uncertainty as to how they
will develop in future. The risk associated with the actuarial modelling infrastructure
36
potentially being unable to identify emerging risks will be picked up as part of a

robust internal systemic risk assessment.
Certain stochastic quantitative approaches may be used to gain insights into past and
emerging sources of external systemic risk. These insights, together with those
gained from the central estimate analysis, will provide useful intelligence on the type
of risks that can emerge in each valuation portfolio, at least the ones that have
emerged in the past. However, one cannot readily assume that past experience is a
reasonable reflection of the future. A more rigorous approach should consider each
of the possible future sources of external systemic risk, using a number of sources of
information.
Communication with business experts
Typically actuaries will hold discussions with portfolio and claim management as part
of the valuation process. These discussions normally provide valuable insights into
emerging trends and possible future sources of external systemic risk. However, the
focus is normally on gaining an appropriate level of portfolio understanding to enable
an informed assessment of the central estimate of outstanding claim and premium
liabilities. Although the information gathered will play a role in the assessment of
risk margins, this tends to be more an afterthought than a key focus of discussions.
Discussions can be readily tailored to topics of relevance for both central estimate and
risk margin purposes and ensure an appropriate level of focus on both aspects of the
valuation process. Business management should be given time to prepare for these
meetings to ensure that the valuation actuary gains the maximum possible benefit
from them.
From a risk margins perspective, the focus of these meetings should be on the
identification of key potential sources of systemic risk, including those that have
begun to emerge and those that may emerge in future. Discussions should consider
all aspects of the portfolio management process, including underwriting and risk
selection, pricing, claims management, expense management, emerging portfolio
trends and the environment within which the portfolio operates. Once the key
sources of external systemic risk have been identified, they can be categorised for
analysis purposes. As well as identifying key risks, the quantification of risk should
be another key consideration for business management interactions.
Selection of assumptions
The selection of CoVs for each risk category will involve a combination of
quantitative analysis and qualitative judgement. Some risk categories will be more
open to quantitative analysis than others. For those categories where such analysis is
more difficult, sensitivity analysis, perhaps in conjunction with business
management, may shed some light on the range of possible outcomes.
In assessing CoVs in respect of each risk category, it is also important to consider the
shape of the entire distribution, to the extent possible. Some risks will demonstrate
characteristics that are reflective of a highly skewed distribution and, as such, may
not have a material bearing on a 75th percentile risk margin but may be more relevant
for higher probabilities of adequacy. An example of such a risk is latent risk where
the probability of such risk emerging is very low and certainly lower than 25%.
Certain sources of superimposed inflation may also be considered to belong to this
category.
37
In focusing efforts, consideration may be given to ranking individual risks for each
valuation class in order of importance, separately for outstanding claim and premium
liabilities. For a number of valuation classes it is quite likely that such an exercise
will identify a small number of key risks and allow efforts to be focused accordingly.
This might also provide justification for excluding certain risk categories that are
deemed to be immaterial in terms of their contribution to the overall CoV. A scoring
system, developed in conjunction with business experts, may be introduced as a
convenient mechanism for ranking individual risks and checking that the
contributions from individual risks to the overall CoV for external systemic risk are
reasonable.
Each risk category will represent the amalgamation of a number of identified

potential sources of risk. In some cases, these individual risks will be correlated and
allowance will need to be made for this when combining the risks to determine a CoV
for the risk category as a whole. A simple approach, similar to that discussed in
section 2.5, may be used to allow for these intra-risk category correlation effects.
A key consideration when determining risk categories for a particular valuation class
is whether there is any correlation between categories. The consolidation of the
analysis of external systemic risk is substantially simplified if one can assume that
each of the risk categories is independent. Certain risk categories may have to be
combined to ensure that this assumption is valid.
In the balance of this section, we explore each of the risk categories discussed in
section 2.5 with a view to providing some insights into the types of risk that may be
included in each risk category and the analysis that may be conducted to estimate
appropriate CoVs for each category.
Economic and social risks
This risk category incorporates a number of potential sources of external systemic

risk. These sources include, but are not limited to, levels of standard inflation (AWE
and CPI), general economic conditions (unemployment rates, GDP growth, interest
rates, asset returns), fuel prices, driving patterns, etc.
Some of these risks can have a material impact on both outstanding claim and
premium liabilities. Others are material only for premium liabilities. For example,
economic conditions can have a material impact on outstanding claim and premium
liabilities for professional lines and builder’s warranty valuation classes. Uncertainty
around driving conditions, on the other hand is less relevant for motor outstanding
claims than it is for motor premium liabilities.
Uncertainty around AWE and/or CPI will impact all valuation classes. Due to the
longer term settlement for long tail classes, AWE uncertainty is somewhat more
material for these classes than for short tail valuation classes. Analysis of past levels
of AWE and CPI can shed some light on past systemic sources of volatility.
Economic commentators often provide insights into the potential sources of volatility.
Any analysis of past levels of inflation should consider the extent to which past
volatility is random and the extent to which it has been impacted by systemic events.
For the purpose of analysis of systemic sources of risk, we are only interested in the
latter. This applies to the analysis of past experience in respect of any systemic event
in any risk category.
Potential systemic shifts in claim frequency for short tail valuation portfolios should
be included in this risk category.
38
Legislative, political and claims inflation risk
These risks have been combined, for convenience, into one risk category since they
are often correlated. For example, the risks associated with the legislative and
political environment are often correlated to the drivers of non-standard claims
inflation for long tail valuation classes.
This risk category is likely to be much more material for long tail valuation classes
than for short tail classes. For long tail classes, in particular, a number of potentially
material risks may be identified and allocated to this risk category. Some of these
risks will be correlated and, as such, quantification should make allowance for this
correlation.
The analysis conducted to quantify CoVs for this risk category can also be used to
justify superimposed inflation assumptions for central estimate valuation purposes.
After all, for long tail valuation classes, the risks in this category are normally
aggregated and referred to as superimposed inflation for insurance liability valuation
purposes. For each risk, one is aiming to form a view of the range of possible
impacts on claim cost outcomes. The average of this range, combined across all
risks, provides an estimate of superimposed inflation.
Individual actuaries will identify the key risks in this category in the context of their
own claims portfolio. As a general guide, for long tail classes, this category would be
expected to include some of the following sub-groups of risk:
Impact of recent legislative amendments, including possibility of erosion of intent

of amendments through assessment and threshold erosion, changes in court
interpretation, etc.
Potential for future legislative amendments with retrospective impacts.
Precedent setting in courts, including impact of judicial decisions perhaps leading
to new heads of damage.
Changes to medical technology costs
Changes to legal costs
Systemic shifts in large claim frequency or severity
Typically, actuaries will have access to various forms of analysis relating to the
potential impact of a specific series of legislative amendments. This information may
include both external and internal analyses, the latter possibly tailored to the specifics
of a particular portfolio. When supplemented by discussions with product and claim
management, a sound understanding of the range of possible outcomes can be
obtained, including the likelihood and potential severity of a particular outcome
occurring.
For short tail classes, this risk category includes the risk that claim inflation will
increase at a level different from that adopted for central estimate purposes, in
addition to that arising from standard inflation (see above) or claim management
process risk (see below). Claim cost reduction initiatives would normally be
allocated to this category and information is sometimes available as to the range of
possible outcomes from such initiatives.
Claim management process change risk
Changes to the claim management process can impact all valuation classes.
Typically, however, such changes will have a more material impact on some
valuation classes than others. The key here is to work closely with claim managers to
39
gain a sound understanding of the claim management philosophy and the process that
underpins that philosophy. Current or future potential changes to process should be
identified as part of such discussions.
Analysis of past experience will help identify past systemic episodes that may have
been impacted by the claim management process. Discussions with management
may help isolate the process changes that contributed to those systemic episodes.
Reporting patterns, payment patterns, finalisation and reopening rates and case
estimation processes should all be considered as part of these discussions.
Sensitivity testing of key valuation assumptions, which can be useful in the

assessment of CoVs for this risk category, is relatively straightforward using
traditional triangulation techniques. If such analysis is conducted, sensitivities
considered should be aligned with the potential sources of uncertainty identified
following discussions with claim management.
Claim management process risk is likely to be more relevant for outstanding claim
liabilities than for premium liabilities. For outstanding claim liabilities, particularly
for short tail valuation classes, this risk can be material since it impacts the pattern of
emergence of credible claim estimates. For premium liabilities, we are more
interested in the extent to which changes to claim management processes can impact
the magnitude of the claim cost. The impact on claim emergence is normally of
secondary importance.
Expense risk
One would generally expect this to be a small contributor to total external systemic
risk.
Ideally, one would spend time with product and claim management to understand the
key drivers of policy maintenance and claim handling expenses. Armed with a good
understanding of these drivers, a valuation actuary can identify the key sources of
possible variation relative to the central estimate assumptions. Sensitivity testing
around the key drivers, preferably conducted in association with informed business
and process experts, and analysis of past expense levels with a view to identifying
past systemic effects can be combined to help form a reasonable view as to the range
of possible claim cost outcomes. Such an analysis could be conducted in conjunction
with any expense analysis conducted for central estimate expense assumptions.
Event claims can have a material impact on the level of claim handling expenses.
The larger an event, the smaller the fixed component of the event management cost
will be as a percentage of the claim cost. In light of this, the analysis may benefit
from including claim handling expenses in respect of event claims with the analysis
of event risk itself.
Event risk
Event risk relates to single events which give rise to a large number of claims. This
risk is likely to be material for property and, to a lesser extent, motor valuation
classes but will be insignificant for most other valuation classes. Event risk also
arises in medical malpractice and builders’ warranty portfolios where a large number
and/or cost of claims can arise from one source, i.e. a single doctor or a single
builder.
The approach to assessing event risk will differ materially between outstanding claim
and premium liabilities. For outstanding claim liabilities, the approach will be
40
defined by the extent to which there are material outstanding events. If there are, then
these should be analysed separately. Discussions with event claim management
should be held to understand their expectations as to claim cost outcomes and to
identify any specific issues that may influence outcomes. The range of development
patterns for previous events may also influence the view on uncertainty.
There is often a wealth of information available to help in the quantification of event

risk for premium liabilities, including:
Past experience in respect of event claims. When analysing past experience, it is

important to allow for changes in portfolio size, geographical spread, inflation,
policy terms and conditions, reinsurance arrangements, etc. where these are
considered to be material. It is not particularly difficult, where sufficient credible
past experience is available, to build a relatively simple statistical model with key
frequency and severity assumptions based on appropriately adjusted past
experience. In fact, modelling of this nature may have been conducted by pricing
actuaries or as part of a reinsurance placement and can be adapted for event risk
analysis.
Output from proprietary catastrophe modelling. A number of such models are
used in practice, including those developed by RMS, EQE, AIR and Risk
Frontiers. Insurers will normally have access to these models through their
reinsurance intermediaries who are well placed to provide advice on the range of
possible outcomes based on modelled events.
Reinsurance intermediaries typically also have available models in respect of
natural perils, and some man-made perils, that can be used to model perils not
covered by proprietary catastrophe models. These, together with the proprietary
models, will normally be used by intermediaries in support of an insurer’s
catastrophe reinsurance program renewal and can be readily extended to provide
advice on the uncertainty associated with event risk.
Latent claim risk
Latent claim risk is negligible for most valuation classes. For some, primarily
workers compensation and liability classes, the risk can be considered to be material.
However, this is one of the most difficult risks to quantify. The probability of these
events is low but the impact should they occur could be substantial
Purely in the context of setting risk margins it is unlikely that analysis of latent claims
risk warrants a substantial commitment of resources given that it is such a low
probability event. However if such risk exposure is significant enough to be a formal
component of the central estimate or if the object of the exercise is modelling extreme
risks for capital adequacy purposes (using a DFA approach) then a thorough
examination of this risk driver is certainly warranted.
.
This risk is the one most likely to be quantified using a large degree of judgement.
Discussions with underwriters may help shed some light on some potential sources
and give a feel for their likelihood and potential impact. Also, casualty reinsurance
underwriters often have a more informed understanding of the potential sources of
latent risk claims from their dealings with a number of direct insurers globally. Using
all of the information collected, scenarios may be developed to reflect a possible
range of scenarios from which reasonable CoVs can be derived.
Recovery risk
This risk category encapsulates systemic uncertainty in relation to reinsurance and

non-reinsurance recoveries. This category is likely to be relatively insignificant for
41
most portfolios. One possible exception is motor valuation classes where third party
recoveries are often a material consideration.
The focus here should be on systemic events that may lead to different recovery
outcomes from those adopted for central estimate purposes.
An analysis of past non-reinsurance recovery rates and patterns will inform on past
systemic events. Combined with discussions with claim management around current
trends in recovery management and any current or planned future initiatives that may
impact recovery levels, one can readily form a view as to the range of possible
systemic outcomes.
Reinsurance recoverability is another potential source of external systemic risk that

should be considered within this category. The extent to which this is material will
depend on the reinsurance arrangements themselves. A material shift in reinsurance
market conditions may significantly alter the ability to recover from reinsurers. For
example, one or more catastrophic events (on a global scale) or a downturn in asset
returns, or a combination of both, may substantially reduce the ability to recover from
reinsurers. The probability of such events occurring and materially impacting
recoveries is low but the severity, should they happen, could be high. Discussions
with reinsurance management are often enlightening and can help in the identification
of possible scenarios, the likelihood of them occurring and the quantitative impact
should they occur.
42
Bibliography
Bateup R and Reed I, 2001, Research and Data Analysis Relevant to the Development of
Standards and Guidelines on Liability Valuation for General Insurance, Proceedings of the
th
13 General Insurance Seminar, Institute of Actuaries of Australia.
th
Collings S and White G, 2001, APRA Risk Margin Analysis¸ Proceedings of the 13 General
Insurance Seminar.
Houltram A, 2003, Reserving Judgement – Considerations Relevant to Insurance Liability

Assessment Under GPS210, Proceedings of the 14th General Insurance Seminar.
Houltram A, 2005, An Empirical Approach to Insurance Liability Predicition Error

Assessment With Application to APRA Risk Margin Determination, Proceedings of the 15th
General Insurance Seminar.
O’Dowd C, Smith A and Hardy P, 2005, A Framework for Estimating Uncertainty in

Insurance Claims Cost, Proceedings of the 15th General Insurance Seminar.
Riley B and Watson B, 2009, Asbestos Liabilities & the New Risk Margins Framework,
Presented to the Institute of Actuaries of Australia 2009 Accident Compensation Seminar.
ROC Working Party, 2008, Reserving Uncertainty, Paper for GIRO 2008
Stochastic Reserving Papers
England P.D. and Verrall R.J., 1999, Analytic and Bootstrap Estimatesof Prediciton Errors in
Claims Reservng, Insurance: Mathematics and Economics 25
England P.D. and Verrall R.J., 2002, Stochastic Claims Reserving in General Insurance,
British Actuarial Journal 8, Part 3
England P.D. and Verrall R.J., 2006, Predicitve Distributions of Outstanding Liabilities in
General Insurance, Annals of Actuarial Science 1, Part 2
Kirschner G.S., Kerley C and Isaacs B, Two Approaches to Calculating Correlated Reserve
Indications Across Multiple Lines of Business
Mack T, 1993, Measuring the Variability of Chain Ladder Reserve Estimates
Taylor G and McGuire G, 2005, Synchronous Bootstrapping of Seemingly Unrelated

Regressions
Verrall R.J., 2004, A Bayesian Generalized Linear Model for the Bornhuetter-Ferguson
Method of Claims Reserving, North American Actuarial Journal, Volume 8, Number 3
43

Marshall - Framework For Assessing Risk Margins

Uploaded by

Copyright:

Available Formats

Marshall - Framework For Assessing Risk Margins

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Marshall - Framework For Assessing Risk Margins

Uploaded by

Copyright:

Available Formats

A Framework for Assessing Risk Margins

A Framework for Assessing Risk

Prepared by the Risk Margins Taskforce

2. The proposed framework ............................................................................... 9

3. Independent risk assessment ............................................................................ 29

4. Systemic risk assessment ............................................................................. 31

The main purpose of this paper is to propose a comprehensive framework for

1.2. Current approaches to assessing risk margins

In preparation for a presentation to the 2006 Reserving Seminar of the Institute of

Although there appear to be a wide range of approaches used by Australian actuaries

Coefficients of variation (CoVs) are determined for individual valuation

More sophisticated approaches include some form of quantitative analysis (stochastic

1.3. Practical framework for assessing risk margins

A number of key stakeholders, including Appointed Actuaries, APRA and auditors,

The proposed framework recognises that quantitative analysis of historical data

In preparing this paper the Taskforce has mainly considered, as a surrounding

1.4. Structure of this paper

In Section 2, we present a framework which takes a methodical and rigorous

Table 1: Summary of risk margin analysis framework

Determine valuation portfolios, claim groups and techniques to deploy

Conduct quantitative analysis, conduct benchmarking where

Apply balanced scorecard approach to objectively score central

Identify, categorise and quantify potential future external sources of

Select correlation coefficients beween valuation classes and between

Consolidate CoVs and correlation coefficients. Independence assumed

Conduct sensitivity testing, scenario testing, internal and external

Document the analysis and judgement relating to each step of the

Conduct annual reviews of key assumptions in the context of emerging

2. The proposed framework

2.1. Introduction to framework

Quantitative techniques alone are insufficient to enable a complete assessment of the

Figure 1: Framework for determining insurance liability risk margins

Valuation Classes represent the

Claim Group is a group of claims

System ic Risk is defined as risks

Risk Risk Frequency Data Error

2.2. Sources of uncertainty

At the highest level, the sources of uncertainty can be categorised as belonging to

Risks internal to the insurance liability valuation process, collectively referred to

In the detailed discussion of the framework below, quantitative and/or qualitative

The nature of traditional quantitative modelling techniques, e.g. bootstrapping and

2.3. Preparing the claims portfolio for analysis

An important consideration is whether the valuation portfolio split adopted to

It may not be possible or particularly insightful, however, to conduct quantitative

2.4. Analysing independent risk sources

Finally, we do consider it useful to supplement any analysis of independent risk for a

2.5. Analysing systemic risk sources

Internal systemic risk

One approach to analysing internal systemic risk is discussed in detail in section 4 of

Development and deployment of a balanced scorecard approach to measuring internal

to draw out all of the risk indicators;

The analysis of internal systemic risk is summarised in Figure 2 below.

Assess specification error using a range of risk

Risk indicators may include number of models

Score each risk indicator against best practice

Assess data error using a range of risk indicators

Risk indicators may include extent, timeliness and

Score each risk indicator against best practice

External Systemic Risk