Alpha and Risk-Adjusted Correlation

for Economic Capital Calculations

Shaun Wang
Department of Risk Management and Insurance,
Robinson College of Business, Georgia State University
March 9, 2007

Abstract: In calculating a firm’s overall capital need and contributions by its

various business lines, the resulting numbers depend heavily upon key risk
assumptions and parameters, including projected profit margin, downside
volatility and correlations. This paper makes a clear distinction between historical
(statistical) correlation in insurance losses versus risk-adjusted correlation in
prices. Firstly, the paper demonstrates that uncertainty in the market price of risk
introduces risk-adjusted price correlation. Secondly, the paper uses risk-adjusted
correlation to reflect sampling errors in the estimation of correlation parameters.
Finally, this paper defines alpha as discounted present value of excess profit
margins in future years, as well as how to explicitly recognize alpha in calculating
economic capitals.

1. INTRODUCTION

Economic capital calculations are at the heart of enterprise-wide risk management of

financial institutions. It is well-known that assumed correlation parameters can have huge
impacts on firm-wide capital requirement and allocated economic capitals to its business
lines. Despite the high importance of such risk parameters, there are alarmingly varying
industry practices regarding correlation assumptions. Consensus on appropriate values of
correlation parameters remains to be elusive. To tackle this issue, this paper makes a clear
distinction between historical versus risk-adjusted correlations. It presents analytical
methods for adjusting correlation parameters to be used for risk aggregation and capital

shaunwang@gsu.edu 1
allocation. This conceptual distinction may help bridge currently diverging industry
practices.

In evaluating risk projects at a given time of specific market environment, the alpha
associated with a risk project refers to expected excess profit margin over some target
benchmark, evaluated over a specified time horizon (say, 1 to 5 years). Treynor and
Black (1973) were pioneers in explicitly reflecting “alpha” in a one-period problem of
optimal asset allocation. Following Treynor and Black’s earlier work, Miller (1999)
advocates using “alpha” in allocating enterprise risk capitals.

For insurance liabilities, there is no actively-traded market from which one can derive
“fair prices”. Most pricing and reserving valuations are performed by marking-to-model.
Market conditions (sentiments) at the time can have a huge impact on the pricing and
reserving valuations. In the property-casualty insurance industry, large deviations from
fair prices can persist for an extended time period, with alternating periods of soft
markets and hard markets. The period of soft market is characterized by lower pricing,
relaxed underwriting standards, more generous coverage provisions; The period of hard
markets is characterized by increased prices, tightened underwriting standards, and
narrowed coverage. This relatively long duration of each soft or hard market makes it
essential to reflect such cyclical market behavior in the economic capital calculations.

2. REVIEW OF INDUSTRY PRACTICES

2.1. Factor-based Regulatory Capital Requirements

Banking and insurance regulations often prescribe capital charges for each type of risk.
For illustration we consider three types of risks {X 1 , X 2 , X 3 } . Let RCj represent the

required capital for risk X j , j= 1, 2, 3.

In the banking Basel Accords, the factor-based capital charges are additive across risks:
RCTotal = RC1 + RC 2 + RC 3 .

shaunwang@gsu.edu 2
In drastic contrast to banking Basel Accords, the U.S. Insurance NAIC RBC (risk-based
capital) explicitly reflects diversification benefit among various types of risks, and
calculates total required capital according to a square-root rule:

RCTotal = RC12 + RC 22 + RC32 .

The different treatments of diversification benefit between the banking Basel Accords
and insurance NAIC RBC lie in their implicit correlation assumptions: The assumed
correlation parameter under Basel Accords is one (or perfect correlation), whereas the
assumed correlation under the insurance NAIC RBC is zero.

Interestingly, such discrepancies also exist among major rating agencies’ approaches. The
S&P factor-based capital charges are additive across risks; while the AM Best BCAR
model has a square-root component as in the NAIC RBC.

2.2. Company Internal Model Approach

Company internal models often represent more advanced approaches than the simple
factor-based capital framework. The internal model approach can be described in three
steps:
Step 1: Individual probability distributions for various types of risks (various
business units, or lines of business) are quantified. By assuming a correlation
matrix, these individual risk distributions are combined together to derive an
aggregate loss distribution for the company over a specified time period.
Step 2: From the aggregate loss distribution one can compute the total company
capital requirement at a prescribed security level, e.g., with 99% probability that
the company will stay solvent over the next time period:
RCTotal = The 99th percentile − Expected Loss.
Remark: supposedly the expected loss shall be covered by revenues (premium or
fees).

shaunwang@gsu.edu 3
 Step 3: The total required capital is then being allocated to various business units
for making risk-based decisions (business planning, product pricing, risk-based
performance measure, etc). See Wang (2002), Ward and Lee (2002).

Both the banking Basel II and EU Insurance Solvency II encourage companies to develop
their own internal risk capital models, as a more advanced approach than the prescribed
factor-based rules. Banking Basel II states that there might be incentives for companies to
build internal models as they are expected to yield a lower required capital than the
factor-based capital requirement. Interestingly, this implicitly admits that although the
factor-based capital requirement is additive across risks, there might be room for
recognizing risk diversification in a more sophisticated internal risk model. In contrast,
the same cannot be said about insurance RBC. Although Insurance RBC is factor-based,
the prevailing practice of using the square-root rule has already given too much credit for
risk diversification. It is doubtful that much more diversification can be derived from
company internal risk models than already included in the square-root rule.

2.3. Capital Allocation

It is well known that correlation parameters can have an important impact on capital
allocation results. In the past few years many insurance companies have launched capital
allocation projects. Using low historical correlation, the capital allocation results often
show large diversification benefits being credited to some lines of business, which have
created much controversy, debate, and confusion among business managers.

Some researchers have advocated various alternatives to using a percentile of the

aggregate loss distribution for setting capital requirement and capital allocation. See
Mildenhall (2003); Venter (2003); Vrieze, and Brehm (2003). A popular risk measure is
the Conditional Tail Expectations (CTE); see Artzner et al (1999). Within the
multivariate normal framework, Panjer (2002) showed that using CTE or percentile will
lead to the same allocation result: the allocated capital to risk X i is proportional to its

relative contribution to the total variance:

shaunwang@gsu.edu 4
 n
σ i ∑ ρ ijσ j
RCi j =1
(2.1) = n
.
∑ρ
RCTotal
σ kσ j
kj
k , j =1

In this paper, our discussion of correlation parameters will be within the multivariate
normal framework (although it allows for straightforward extensions to normal copulas).
As a result, using either percentile or CTE would not make any material difference for
capital allocation discussions.

3. HISTORICAL VERSUS RISK-ADJUSTED CORRELATION

In this paper, I propose to make a clear distinction between historical (observed) versus
risk-adjusted correlation. To help articulate this conceptual distinction, here I draw an
analogy in bond default risk probabilities.

Historical versus Implied Default Probabilities

In modeling corporate bond defaults, it is widely known that historical default probability
can differ dramatically from the implied default probabilities as implied from bond yield
spreads. As shown in Table 3.1 and Figure 3.1, the default probabilities implied by yield-
spreads are much higher than historical default probabilities. Their relative differences
are most striking for high-grade bonds (with low expected default frequencies). Such
differences represent the market’s explicit adjustments for risk, uncertainty and illiquidity.

Table 3.1. Annualized Average Credit Loss Rates for 5-year bond (1980-2004) v.s.
Bloomberg zero coupon spreads to Aaa, (ref: Berndt et al, 2005).
rating yield spread default rate
Aa 0.0009 0.0002
A 0.0022 0.0006
Baa 0.0100 0.0027
Ba 0.0166 0.0143

shaunwang@gsu.edu 5
 Figure 3.1 Annualized Average Credit Loss Rates for 5-year bond (1980-2004) v.s.
Bloomberg zero coupon spreads to Aaa. (ref: Berndt et al, 2005)

Annualized Average Credit Loss Rates

(1980-2004)

0.02

yield spread
Default Rate

0.015
default rate
0.01

0.005

0
Aa A Baa Ba
Bond Rating

Setting a capital charge for a risk is analogous to assigning a risk-adjusted price for the
risk. In the same way that implied default probabilities differ from historical default
probabilities, conceptually, we should expect to use risk-adjusted correlation (rather than
historical correlation) in allocating economic capitals.

For insurance companies, a straightforward statistical analysis of insurance losses may

indicate very low correlation among different lines of business. This was the basis for the
square-root rule in the insurance RBC approach. However, upon a deeper analysis of the
nature of insurance business, for capital allocation purposes we have reasons for using
higher risk-adjusted correlations than historical (statistical) correlation. For capital
allocation purposes, to reflect the inherent risks and illiquidity of insurance contracts, the
large transaction costs associated with maintaining or running-off the insurance
operations, we should use risk-adjusted correlations that are higher than historical
correlations.

shaunwang@gsu.edu 6
4. RISK-ADJUSTED PRICE CORRELATION

Consider an insruance risk X and a reference portfolio Z. Here X can represent the losses
from a product or business line; Z represents losses from a reference portfolio (the
company portfolio, or the industry portfolio).

Firstly, with the passage of time, k =0, 1, 2, 3, …, once new information becomes
available, we update our probabilistic estimates for the loss variables X and Z:
(4.1) μˆ X (k ) = μ X0 (k ) + ε X (k ) , and μˆ Z (k ) = μ Z0 (k ) + ε Z (k ) .
Here μ X0 (k ) represents the true mean value of the statistical distribution of the
prospective loss variable X, and ε X (k ) represents the estimation error contained in the
estimated mean loss amount μˆ X (k ) . For the estimation error terms, we denote

σ X2 (k ) = Var (ε X (k ) ) ,
σ Z2 (k ) = Var (ε Z (k ) ) , and
ρ X ,Z (k ) = Corr ε X (k ), ε Z (k ) .
Next, we employ pricing methods to derive fair risk-adjusted prices for X and Z.

Assume that the prices for X and Z are obtained through the following equations:
(4.2) hX (k ) = μ X0 (k ) + ε X (k ) + λ (k ) ⋅ σ X (k ) , and

hZ (k ) = μ Z0 (k ) + ε Z (k ) + λ (k ) ⋅ σ Z (k ) .
The quantity λ(k) is called the market price of risk, which may well change over time as
market condition changes. At any specific time, the same market price of risk is used to
derive the prices of X and Z simultaneously. At time t = k, the uncertainty in the market
price of risk is measured in terms of variance of λ:
σ λ2 (k ) = VaR(λ (k ) ) .
In property-casualty insurance, we can estimate σλ over a time period that spans over at
least one underwriting/pricing cycle.

shaunwang@gsu.edu 7
For the sake of simplicity, we assume zero correlation between the estimation error
ε Z (k ) , and the market price of risk, λ (k ) , that is
(4.3) Cov ε X (k ), λ (k ) = 0 , and Cov ε Z (k ), λ (k ) = 0 .

Later on we can relax this constraint especially in discussing natural catastrophic

insurance.
We can now derive the correlations between their risk-adjusted prices:
( )
VaR hX (k ) = σ X2 (k ) ⋅ 1 + σ λ2 (k ) .

VaR h (k ) = σ (k ) ⋅ (1 + σ (k ) ).
Z
2
Z
2
λ

Cov hX (k ), h (k ) = σ (k ) ⋅ σ (k ) ⋅ (ρ (k ) + σ
Z X Z X ,Z
2
λ )
(k ) .

Thus we obtain the following correlation between the “prices”:

ρ X , Z (k ) + σ λ2 (k )
(4.4) ρ *
(k ) = Corr hX (k ), hZ (k ) = .
1 + σ λ2 (k )
X ,Z

One can verify that ρ *X , Z (k ) ≥ ρ X , Z (k ) . In other words, uncertainty regarding the market

price of risk induces a higher correlation in prices than the loss correlation. Note that
σ λ2 (k )
1. when ρ X , Z (k ) = 0 we have ρ X* ,Z (k ) = > 0. Even though business lines
1 + σ λ2 (k )
may appear to have low statistical correlation in terms of losses, uncertainty in the
market price of risk may induce higher positive correlation in the “risk-adjusted”
prices.
2. when ρ X , Z (k ) = 1 we have ρ *X , Z (k ) = 1

σ λ2 (k ) − 1
3. when ρ X , Z (k ) = −1 we have ρ *X ,Z (k ) = > −1 .
σ λ2 (k ) + 1

shaunwang@gsu.edu 8
Figure 4.1 Uncertainty in “Market Price of Risk” and “Adjusted Correlation”

Mapping rho to rho*

when sigma(lambda)=0.5
1

0.8

0.6

0.4

0.2
rho*

0
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-0.2

-0.4

-0.6

-0.8

-1
rho

5. CONFIDENCE INTERVALS FOR CORRELATION PARAMETERS

Given a sample of “n” pairs of observations: {(x j , y j ), j = 1,Λ , n}, we can calculate

the sample correlation coefficient:

n n ⎛ n ⎞⎛ n ⎞
∑ (x j − x )( y j − y ) n ⋅ ∑ x j y j −⎜⎜ ∑ x j ⎟⎟⎜⎜ ∑ y j ⎟⎟
(5.1) rxy =
j =1
=
j =1 ⎝ j =1 ⎠⎝ j =1 ⎠ ,
(n − 1) ⋅ s x ⋅ s y n ⎛ ⎞
n
2
⎛
n ⎞ n
2

n∑ x 2j − ⎜⎜ ∑ x j ⎟⎟ ⋅ n∑ y 2j − ⎜⎜ ∑ y j ⎟⎟
j =1 ⎝ j =1 ⎠ j =1 ⎝ j =1 ⎠
where x and y are sample means of sx and sy are sample standard deviations.

In statistics, hypotheses about the value of the population correlation coefficient ρ

between variables X and Y of the underlying population, can be tested using the Fisher
transformation applied to the sample correlation r.
1 1 + rxy
(5.2) z xy = FISHER(rxy ) = ln ,
2 1 − rxy

which is called Fisher transformation.

shaunwang@gsu.edu 9
If (X, Y) has a bivariate normal distribution with correlation coefficient ρ, then the Fisher
transform (5.2) is approximately normally distributed with mean
1 1+ ρ
(5.3) ln
2 1− ρ

and standard deviation 1 n − 3 . This property forms the basis for a common way of
constructing a confidence interval for ρ. For the Pearson correlation coefficient ρ, the
100α% confidence interval is

⎛ z1−α ⎞ ⎛ z1−α ⎞
(5.4) FISHER −1 ⎜ z xy − 2 ⎟
, FISHER −1 ⎜ z xy + 2 ⎟
,
⎜ n − 3 ⎟ ⎜ n − 3 ⎟
⎝ ⎠ ⎝ ⎠
where
¾ zij is the Fisher z-transform of the correlation coefficient:
⎛ α⎞
¾ z1−α = Φ −1 ⎜1 − ⎟ is the 100(1-α/2) percentile of the standard normal variable;
2 ⎝ 2⎠
¾ The inverse Fisher transform is:
sinh( x) e z − e − z e 2 z − 1
(5.5) FISHER −1 ( x) = tanh( z ) = = = .
cosh( x) e z + e − z e 2 z + 1

Table 5.1. For a sample size of 15, the 95% confidence interval for estimated historical
correlation

Corr Lower Corr* Upper Corr*

(1.00) (1.00) (1.00)
(0.80) (0.93) (0.49)
(0.60) (0.85) (0.13)
(0.40) (0.76) 0.14
(0.20) (0.65) 0.35
0.00 (0.51) 0.51
0.20 (0.35) 0.65
0.40 (0.14) 0.76
0.60 0.13 0.85
0.80 0.49 0.93
1.00 1.00 1.00

shaunwang@gsu.edu 10
Interestingly, even if the sample correlation coefficient is zero, the 95% confidence
interval upper limit still indicates a strong “0.51”! The large range of feasible “ρ”
cautions us not to put too much confidence in historical correlation parameters, and even
calls for an explicit adjustment of the historical correlation.

Based on the Fisher transform for constructing confidence intervals for the correlation
parameter, we give the following mathematical formula for making an explicit
adjustment in the correlation parameter for sampling errors:
⎛1 1+ ρ 1 ⎛ α ⎞⎞
(5.6) ρ * = tanh⎜⎜ ln + ⋅ Φ −1 ⎜1 − ⎟ ⎟⎟ or
⎝2 1− ρ n−3 ⎝ 2 ⎠⎠

⎛ 1 ⎛ α ⎞⎞
ρ * = FISHER −1 ⎜⎜ FISHER( ρ ) + ⋅ Φ −1 ⎜1 − ⎟ ⎟⎟ .
⎝ n−3 ⎝ 2 ⎠⎠
The ρ* in (5.6) adjusts for estimation error in the correlation coefficient. This adjustment
is especially important when the correlation parameter is estimated from small sample
data, as it is the case in property-casualty insurance whereas we have mostly 10 to 20
years of historical observations.

Figure 5.1
Correlation Confidence Intervals
95% Confidence Interval, sample size=15
1

0.8 rho

0.6 upper rho*

0.4 lower rho*

0.2

0
-1

-0.8

-0.6

-0.4

-0.2

0.2

0.4

0.6

0.8

-0.2

-0.4

-0.6

-0.8

-1

shaunwang@gsu.edu 11
Figure 5.2

Mapping rho to upper rho*

75% confidence interval, sample size = 20
1.0

0.8

0.5

0.3
upper rho*

0.0
-1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0
-0.3

-0.5

-0.8

-1.0
rho

Remark 5.1. When more than two risks are involved, one should check for the positive
definite property after adjusting the correlation parameters.

⎛ 1 ρ12 ρ13 ⎞ ⎛ 1 ρ12* ρ13* ⎞

⎜ ⎟ ⎜ ⎟
⎜ ρ12 1 ρ 23 ⎟ Æ ⎜ ρ12* 1 ρ 23
*
⎟
⎜ρ ρ 23 1 ⎟⎠ ⎜ ρ* ρ 23
*
1 ⎟⎠
⎝ 13 ⎝ 13

6. IMPLICATIONS IN CAPITAL ALLOCATION

Recall that within the multivariate normal framework, the capital allocation to risk X i is a
covariance-based allocation; see equation (2.1). Using risk-adjusted correlation
parameters, the covariance allocation method yields a different allocation:
n
σ i ∑ ρ ij*σ j
RCi* j =1
*
= n
.
∑ρ
RCTotal *
kjσ kσ j
k , j =1

Example 6.1 Consider three risks X1, X2 and X3 with

shaunwang@gsu.edu 12
 ⎛ 1 ρ12 ρ13 ⎞ ⎛ 1 0 0 ⎞
⎜ ⎟ ⎜ ⎟
(σ 1 , σ 2 , σ 3 ) = (1, 1, 8) , and ⎜ ρ12 1 ρ 23 ⎟ = ⎜ 0 1 0 ⎟ .
⎜ρ ρ 23 1 ⎟⎠ ⎜⎝ 0 0 1 ⎟⎠
⎝ 13

(1) Using different volatilities for the market price of risk, we get different risk
adjusted correlations, which in turn yield different capital allocation results:

Attribution σ(λ)=0 σ(λ)=0.5 σ(λ)=1

to X1 1.5% 3.8% 6.6%
to X2 1.5% 3.8% 6.6%
to X3 97.0% 92.4% 86.7%

Note that σ(λ)=0 corresponds to the case of using historical correlation without
adjustment; the proportions of allocated capitals to risks X1 and X2 are very small.
As the volatility of λ increases, the proportions of allocated capital to risks X1 and
X2 also increase.

(2) Suppose that the correlation parameters are estimated from 20 observations,
and use the upper bound of the 75% confidence interval for the estimated
correlation parameters, we get the following allocated capitals:

Upper bound of 75%

Attribution Unadjusted Correlation confidence interval
to X1 1.5% 4.6%
to X2 1.5% 4.6%
to X3 97.0% 90.8%

shaunwang@gsu.edu 13
7. PROJECTING ALPHA IN UNDERWRITING/PRICING CYCLE

Historically, the market price of risk over time exhibited cyclical behavior, rather than a
random pattern. We use a simplistic model for the cyclical profit margins as
⎛ j ⎞
(7.1) PM j = A ⋅ sin ⎜ ⎟, j = 0,1,2,…
⎝ 2bπ ⎠
where “b” is the length of a complete underwriting cycle, and “A” is the amplitude of the
cycle. We can calibrate the parameter “A” so that the underwriting year excess profits
exhibit a volatility that matches the observed underwriting year loss ratio volatility.

Discounted Cash Flow Formula for Calculating “Alpha”

Recall that risky assets can be valued according to the Discounted Cash Flow equation:
N FV j
DPV = ∑ ,
j =1 (1 + R) j
where
• FVj is the nominal value of cash flow amount at the end of future period (j−1, j];
• R is the interest rate for discounting

Applying the Discounted Cash Flow valuation approach, we define the prospective
“alpha” as “Discounted Excess Profit Margin”:
N EPM j
(7.2) α =∑
j =1 (1 + R) j
where
o “N” is the maximum number of future years to be considered (for instance “N”
can be between 2 and 5 years).
o EPMj is the projected profit margin at the end of future period (j−1, j];
o R is the interest rate for discounting, which should reflect the level of uncertainty
in the projected profit margins.

Fair Profit Margin Implies A Benchmark Amount of Economic Capital

shaunwang@gsu.edu 14
For risk X and reference portfolio Z. Assume that the market prize of risk for the
reference portfolio Z is λ0Z . Based on CAPM-type of approach, we can derive a market
price risk for risk X as:
λ0X = λ0M ⋅ ρ *X .
This reinforces the importance of the risk-adjusted correlation parameter ρ *X . Base on

previous discussions, there are multiple justifications for using a value of ρ *X that is
significantly greater than zero (or even close to one).

The fair profit margin for risk X is then λ0X ⋅ σ , which in turn implies an amount of
economic capital for taking risk X :
λ0X ⋅ σ X (k )
(7.4) EC X (k ) = ,
TEROE
where
¾ TEROE is a target excess rate of return (over the risk free rate). For instance, we
may assume that TEROR = 10%.
¾ λ0X is the long-term target average market price of risk for the given line of
business, which has already reflected the risk-adjusted correlation with reference
portfolios. For instance, we may assume that λ0X = 0.3.

The economic capital assigned to risk X, when adjusted for “alpha”, should be
λ0X ⋅ σ X (k )
(7.5) EC X (k ) = −α .
TEROE

Example 7.1: Assume that Underwriting Year loss ratio volatility for commercial
auto liability is σ=6%; after adjusting for multi-year developments the annualized
volatility is 4.2%. We also assume a simple cyclical model for the excess profit
margin as
⎛ j ⎞
EPM j = A ⋅ sin ⎜ ⎟, j = 0, 1, 2, …
⎝ 2bπ ⎠

shaunwang@gsu.edu 15
 where “b=10 years” is the length of a complete underwriting cycle. We calibrate the
parameter “A=0.0855” so that the underwriting year profit margins exhibit a volatility
of σ=6%.
We then use a Discounted Cash Flow formula to calculate alpha value in year k:
j
5
⎛ 1 ⎞
α (k ) = ∑ ⎜ ⎟ ⋅ EPM k + j , for k = 0, 1, 2, …
j =1 ⎝ 1 + R ⎠

Assume that β =1, TEROR = 10%, R=100%, and λ0X = 0.3, we get the following:

⎡ 0.3 ⎤
EC = ⎢ ⎥ × 4.2% = 0.126 .
⎣ 0.1 ⎦
Before reflecting alpha value, the “Economic Capital Factor” is 0.126 (to be applied
to the expected loss portion of the premium).
After adjusting for alpha, the ‘Economic Capital Factors” would depend on the
current phase of the underwriting cycle:
⎡ 0.3 ⎤
EC (k ) = ⎢ ⎥ × 4.2% − α (k ) = 0.126 − α (k )
⎣ 0.1 ⎦

time k α (k) EC(k)

0 0.029 0.097
1 0.059 0.067
2 0.066 0.060
3 0.048 0.078
4 0.011 0.115
5 (0.029) 0.155
6 (0.059) 0.185
7 (0.066) 0.192
8 (0.048) 0.174
9 (0.011) 0.137
10 0.029 0.097

shaunwang@gsu.edu 16
 Figure 7.1
0.1
Alpha
0.08
Phase of Cycle
0.06

0.04

0.02

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
-0.02

-0.04

-0.06

-0.08

-0.1

Figure 7.2
Projected Economic Capital Factors

0.25

alpha(k)
0.20
EC(k)
0.15

0.10

0.05

0.00
1 2 3 4 5 6 7 8 9 10 11
-0.05

-0.10
Year

Remark 7.1: In real life, year over year changes in profit margin do not completely
follow a random generator, but rather like a slow motion display of the roller coast of
cycle swings. This characteristic makes it plausible to estimate “alpha” in a timely
manner. This highlights the importance of explicitly incorporating “alpha” in the
economic capital calculations for property-casualty insurers.

shaunwang@gsu.edu 17
8. A REAL-LIFE CASE OF NATURAL CATASTROPHE INSURANCE

Natural catastrophe losses, by definition, are low frequency and high severity events.
A catastrophic loss such as Hurricane Katrina in year 2005 can simultaneous impact the
estimated mean loss and the market price of risk in subsequent years.
Recall that
hX (k ) = μ X0 (k ) + λ (k ) ⋅ σ X (k ) + ε X (k ) .
Immediately following a major catastrophic loss in year k, both the revised estimate of
the mean μˆ X (k + 1) , and the market price of risk, λ (k + 1) , would jump upwardly
simultaneously:
Cov ε X (k + 1), λ (k + 1) > 0 and Cov ε Z (k + 1), λ (k + 1) > 0.

This is exactly the time that there is significantly positive “alpha” in the increased
insurance prices. Explicit recognition of “alpha” would offset (to some extent) the
increased capital requirement for insurers continue to provide catastrophic insurance
coverage.

Here we have a real-life drama unfold post the 2005 Hurricane Katrina, which marked
the largest ever natural catastrophic losses in the U.S. history (up to the time of writing).
As aftermath of Hurricane Katrina, the insurance industry experienced a major capacity
crunch, and insurance prices in Gulf coast areas increased significantly for reduced limit
coverage. To make things worse, Risk Management Solutions (a major commercial
catastrophe modeling firm) released a revised CAT Model version 6.0 which significantly
raised Possible Maximum Loss (PML) estimates. Borrowing a description by Don Mango,
“the increased PMLs had ripple effects, from more stringent rating agency stress tests to
put additional pressure on reinsurer capacity constraints, to cedents looking to buy more
limit, and to prices going up.” A discontinuous market shock has had detrimental effects
across the insurance industry and adversely affected the general public and business
community in need of catastrophe insurance coverage. If rating agencies gave due
consideration of “positive alpha” contained in the 2006 high-flying insurance prices, it

shaunwang@gsu.edu 18
would have alleviated some of the “capacity crunch” and would have helped dampen the
severity of price jumps.

9. AREAS FOR FURTHER EXTENSIONS AND RESEARCH

In this paper we make a clear distinction between historical correlation and risk-adjusted
correlation. Conceptually this is analogous to the distinction between historical default
probability versus implied default probability. We demonstrated that uncertainty in the
market price of risk have an effect of increasing correlation in the prices. We also
discussed the sampling errors involved in the estimation of correlation parameters and
construction of confidence intervals for the estimated correlation parameter. This paper
presents analytical methods for transforming historical correlation to risk-adjusted
correlation. The insurance RBC square-root rule might have given too much
diversification benefits. In contrast, the banking Basel II additive capital charges do not
allow for diversification benefits. By using risk-adjusted correlations, we can find a
middle ground that better reflects the true risk contributions from individual risks.

This paper advocates explicit recognition of “alpha” in economic capital calculations.

The benefits of making explicit adjustment of alpha include dampening the amplitude of
underwriting cycle. The effectiveness of using “alpha” depends upon how timely and
accurately we can measure the alpha in the lines of business.

One area of further research is to investigate the implications of using risk-adjusted

correlation when using other capital allocation methods, such as those discussed in
Phillips, Cummins and Allen (1998), Meyers & Read (2001), Sherris (2004), and others.

shaunwang@gsu.edu 19
REFERENCES:

Artzner, P., F. Delbaen, J. M. Eber, and D. Heath. 1999. Coherent measures of risk.
Mathematical Finance 9 (November): 203-228.

Berndt, A., Duffie, D., Ferguson, M., Douglas, R., Schranz, D. (2005) Measuring
default-risk premium from default-swap rates and EDFs, presentation at the Moodys-LBS
Credit Risk Conferences, London, 2005.

Bodie, Zvi, Kane, Alex, and Alan Marcus, Investments, Irwin McGraw-Hill, 3rd edition,
pp. 760-769.

Malkiel, B., 1973, A Random Walk Down Wall Street, W. W. Norton and Company.

Markowitz, H., 1959, Portfolio Selection: Efficient Diversification of Investments, John

Wiley & Sons.

Mildenhall, S. (2003) A Note on the Myers and Read Capital Allocation Formula,
Casualty Actuarial Society Forum, Fall 2003, 419-450.

Miller, Ross M., 1999, Treynor-Black Revisited: A New Application to Enterprise-Wide

Portfolio Optimization
http://home.earthlink.net/~millerrisk/Papers/TreynorBlackRevisited.htm

Myers, S. and Read, J. (2001) Capital Allocation for Insurance Companies, Journal of
Risk and Insurance, 68(4), 597-636.

Panjer, H.H. (2002) Measurement of Risk, Solvency Requirements and Allocation of

Capital within Financial Conglomerates, AFIR, Cancun, 2002.

Phillips, R., Cummins, D, and Allen, F. (1998). Financial Pricing of Insurance in the
Multiple Line Insurance Company, Journal of Risk and Insurance, 65(4), 597-636.

Sherris, M. (2004) Solvency, Capital Allocation and Fair Rate of Return in Insurance,
preprint. Available from the CAS website:
http://www.casact.org/research/summaries/allocation.htm

Treynor, J. L. and F. Black, 1973, How to Use Security Analysis to Improve Portfolio
Selection, Journal of Business, January, pages 66-88.

Venter, G. (2003) Discussion of Myers and Read “Capital Allocation for Insurance
Companies,” Casualty Actuarial Society Forum, Fall 2003, 459-478.

Vrieze, K. and Brehm, P. (2003) Review of Myers and Read “Capital Allocation for
Insurance Companies,” Casualty Actuarial Society Forum, Fall 2003, 479-491.

shaunwang@gsu.edu 20
Wang, S. (1996). “Premium Calculation by Transforming the Layer Premium Density.”
ASTIN Bulletin, 26 (1996): 71-92.

Wang, S. (2000). “A Class of Distortion Operators for Pricing Financial and Insurance
Risks.” Journal of Risk and Insurance, 67 (2000 March): 15-36.

Wang, S. (2002). A Set of New Methods and Tools for Enterprise Risk Capital
Management and Portfolio Optimization, Casualty Actuarial Society Forum, Summer
2002. http://www.casact.org/pubs/forum/02sforum/02sftoc.htm

Wang, S. and Faber, R. (2006). Enterprise Risk Management for Property-Casualty

Insurance Companies, ERM Institute International research report,
http://www.ermii.org/Research/downloads/erm_paper080106.pdf

Ward, L., and Lee, D. (2002). Practical Application of the Risk-adjusted Return on
Capital Framework, Casualty Actuarial Society Forum, summer 2002.
http://www.casact.org/pubs/forum/02sforum/02sftoc.htm

shaunwang@gsu.edu 21