foundations of quantitative risk measurement€¦ · f hedging the liabilities using a replicating...
TRANSCRIPT
Foundations of Quantitative Risk Measurement
Chapter 6: Risk measures
Jan Dhaene and Daniel Linders
November, 2019
0 – Outline 2/41
1. Introduction
2. Risk measuresDefinitionCoherent risk measures
3. Value-at-Risk and Tail Value-at-Risk
4. The required solvency capital
5. The Haezendonck-Goovaerts risk measure
6. VaR, TVaR and comonotonicity
1 – The insurance business 3/41
Inverted production cycle:I Insured: pays a fixed premium.
I Insurer: accept the risk to pay the claim amounts related to possiblefuture events.
I Similar for insurer/reinsurer or reinsurer/reinsurer.
Solvency:I The insurer has to set a premium, such that he is able to pay all the
future benefits to the policy holders.
I When a claim has to be paid, the insurer should be solvent.
Example:I X and/or T can be random.
1 – Risks for an insurer 4/41
Underwriting risk:I The risk that the premiums are not sufficient to cover the obligations
of the insurer.
I Underwriting risk arises when the insurer underestimates the futurelosses.
Credit risk:I The risk that a counter party will not meet his obligations.
I The premiums are invested in bonds (and/or other assets), which candefault.
Market risk:I Exposure to the uncertain future market value of the investment
portfolio.
Other risks:I Liquidity risk, operational risk, . . .
1 – Solvency 5/41
Technical provisions
Technical provisions:I also called: actuarial reserves or best estimate;
I the amount the insurer has to hold in order to cover the expectedfuture claims;
I fair value of the future liabilities: the amount another party is willing topay to take over the insurance business.
Valuation = Market ConsistentI Mark-to-market:
F hedging the liabilities using a replicating portfolio gives a unique price.
I Mark-to-modelF Model-based approach to determine the expectation of the future
liabilities.
F Model risk: use a prudence margin,
1 – Solvency 6/41
What is solvency:I Solvency refers to the ability of the insurer to meet his obligations to
pay the present and future claims related to policyholders.
Solvency rules:I Imposed by the regulatory authority.
I Protecting policyholders.
Solvency Capital Requirements:I In addition to the technical provisions V, the insurer has to hold a
capital buffer.
I Minimal level of capital an insurer has to hold such that the insurer isvery likely to be able to meet his future obligations.
1 – Solvency II 7/41
http://www.solvency-2.com
2 – Risk measures 8/41
Definition
Definition (Risk measure)
Consider a set Λ of real-valued r.v.’s. The function ρ assigning a realnumber ρ [X] to a r.v. X ∈ Λ, is called a risk measure.
ρ is also called a risk measure with domain Λ.I In most cases, Λ is not specified and has to be taken ‘as broad as
possible’.
X represents a loss over a given reference periodI ρ [X] captures the ‘risk’ in a real number.
Example:I Standard deviation risk measure
I ρ [X] = Expected value + safety loading:
ρ [X] = µX + λσX, λ ≥ 0.
2 – Risk measures 9/41
References
Risk measures and premium principles:I Goovaerts, De Vylder & Haezendonck (1984), Kaas, Goovaerts,
Dhaene & Denuit (2008).
Risk measures and capital requirements:I Artzner (1999), Wirch & Hardy (2000).
2 – Properties 10/41
Distributional properties that a risk measure may satisfy
Law invariance:
X d= Y⇒ ρ [X] = ρ [Y] .
Preserving stop-loss order:
X sl Y⇒ ρ [X] ≤ ρ [Y] .
Comonotonic additivity:I For any comonotonic r.v.’s X, Y:
ρ [X + Y] = ρ [X] + ρ [Y] .
2 – Example: the stop-loss risk measure 11/41
Stop-loss premium with retention K ∈ R :I The risk measure ρ is defined as
ρ [X] = E[(X− K)+
].
ExerciseI Prove that ρ is law invariant.
I Prove that ρ preserves stochastic dominance and stop-loss order.
Prove also that ρ is not comonotonic additive.
2 – Properties 12/41
Distribution-free properties that a risk measure may satisfy
Monotonicity:
X ≤ Y⇒ ρ [X] ≤ ρ [Y] .
Positive homogeneity:I For any a > 0,
ρ [aX] = aρ [X] .
Translation invariance:I For any b ∈ R
ρ [X + b] = ρ [X] + b.
Subadditivity:
ρ [X + Y] ≤ ρ [X] + ρ [Y] .
2 – Coherent risk measures1 13/41
Definition (Coherent risk measure)
A risk measure is said to be coherent if it satisfies the properties
1. monotonicity;
2. positive homogeneity;
3. translation invariance;
4. subadditivity.
1Artzner et al. (1999)
2 – Upper expectation 14/41
Example of a coherent risk measure
Upper expectation2
I Π is a subset of all probability measures on the measurable space(Ω,F )
ρΠ [X] = sup EP [X] | P ∈ Π .
I Exercise: Prove that ρΠ is a coherent risk measure.
Interpretation:I The elements of Π are generalized scenarios.
I Expectation of X with respect to a worst-case scenario.
Upper expectations can also be considered as a G-expectations.3
2see Huber (1981)3see e.g. Peng (2007)
2 – Upper expectation 15/41
Law invariance and upper expectation
Define the risk measure ρ as:
ρ [X] = max E [X] , E [X | Z > z] ,
with P [Z ≤ z] > 0.
ρ is an upper expectation:
I P is the real-world probability measure.
I Q is a distorted probability measure:
Q [A] = P [A | Z > z] , for an event A.
I Then:ρ [X] = max
EP [X] , EQ [X]
.
ρ is not law invariant:
I For Y d= Z and Y and Z independent: ρ [Y] 6= ρ [Z] .
2 – Upper expectation 16/41
Risk and uncertainty
In Knight (1921), Frank Knight makes a distinction between risk anduncertainty.
Risk refers to the uncertainty about the realization of a loss X.I Risk is modeled by the cdf FX of X, which gives the probability of X
being in a given region.
I If the cdf FX is known, the expectation E[X] can be determined.
In practical situations, the cdf FX is not fully known.I We only have partial information: FX ∈ P .
I This induces an extra source of uncertainty.
A robust estimate for the expectation is the upper expectation
sup EF [X] | F ∈ P .
2 – Coherent risk measures and upper expectation 17/41
Theorem
Consider a risk measure ρ. The following statements are equivalent:
1. The risk measure ρ is coherent.
2. There exist a set Π of probability measures such that
ρ [X] = sup EP [X] | P ∈ Π .
A proof for the case where Ω is finite: Huber (1981).
The more general case is proven in Delbaen (2002).
3 – Value-at-Risk 18/41
Definition
Definition
For any p in (0, 1) the Value-at-Risk at level p is defined by
VaRp [X] = inf x ∈ R | FX (x) ≥ p .
Value-at-Risk = Quantile
I VaRp [X] = F−1X (p)
Interpretation:I Take p close to 1.
I Probability that X exceeds the threshold is small:
P[X > VaRp [X]
]= 1− FX
(VaRp [X]
)︸ ︷︷ ︸≥p
≤ 1− p.
3 – Properties of Value-at-Risk 19/41
Value-at-Risk is increasing in pI p < q⇒ VaRp [X] ≤ VaRq [X] .
Positive homogeneous and translation invariant:I For a > 0 and b ∈ R : VaRp [aX + b] = aVaRp [X] + b.
I Exercise: Prove these properties.
VaR is not a coherent risk measureI Exercise: give a counterexample.
3 – Solvency capital requirement 20/41
The VaR for avoiding bankruptcy
Solvency capital requirement:I X = future loss/liabilities.
I V = technical provision (fair valuation of X).
I K [X] = solvency capital
K [X] = VaRp [X]−V
I Extra buffer of capital for bad times4.
Interpretation:I K [X] =VaRp [X−V] .
I X−V = loss which is not covered by the technical provisions
P [X−V > K [X]] = P [X > K [X] + V] ≤ 1− p.
I Insolvency = technical provision + capital buffer are not sufficient tocover to loss:
F For p close to 1, the probability that the insurer is insolvent is small.4p should be sufficiently big, such that K [X] > 0.
3 – Example 21/41
VaR for normal and lognormal distributions
The cdf of a standard normal distribution = Φ.
Assume Y d= N
(µ, σ2) . Then
VaRp [Y] = µ + σΦ−1 (p) .
For a lognormal r.v. X d= eY :
VaRp [X] = eµ+σΦ−1(p).
3 – Problems with the Value-at-Risk 22/41
VaR and low frequency - high impact events
Consider the r.v. X:
P [X = −10000] = 0.999 and P [X = 10 000 000] = 0.001.
high probability of having a gain, but a low probability of having a bigloss.
The VaR does not detect there is an extreme risk:
VaR0.95[X] = −10000.
However, if something goes wrong, i.e. if the loss exceeds the VaR,the loss is enormous:
E [X | X > VaR0.95[X]] = 10 000 000.
3 – Value-at-risk 23/41
“I believe that VaR is the alibi that bankers will give shareholders(and the bailing-out taxpayer) to show documented due diligence, andwill express that their blow-up came from truly unforeseeablecircumstances and events with low probability not from taking largerisks that they didn’t understand. I maintain that VaR encouragesuntrained people to take misdirected risks with shareholders’, andultimately the taxpayers’, money.”
———- Nassim Taleb 1997.
“The risk-taking model that emboldened Wall Street to trade withimpunity is broken and everyone is coming to the realization that noalgorithm can substitute for old-fashioned due diligence. VaR failed todetect the scope of the market’s collapse. The past months haveexposed the flaws of a financial measure based on historical prices.”
———- Financial Reporter Christine Harper, January 2008.
3 – Tail Value-at-risk 24/41
Definition
For any p in (0, 1) the Tail Value-at-Risk at level p is defined by
TVaRp [X] =1
1− p
∫ 1
pVaRq [X] dq.
Interpretation:I TVaR is an average of VaR’s.
I Exercise: prove that VaRp [X] ≤ TVaRp [X] .
What is bad?I Probability that X exceeds VaRp [X] is small, but not 0.
I If X exceeds VaRp [X] , what will be the loss?
I Bad times are those when X ∈[VaRp [X] , TVaRp [X]
].
3 – Example 25/41
TVaR as a cushion for extreme losses
Consider the loss X :
P [X = 0] = 0.6,P [X = 100] = 0.37,
P [X = 10000] = 0.02,P [X = 100000] = 0.01.
VaR0.95 [X] = 100.I Probabillity of a loss bigger than 100 is ‘only’ 5%.
I What can we expect if the loss will exceed VaR0.95 [X] ?
TVaR0.95 [X] = 24040.I If the loss will exceed VaR0.95 [X] , then it will be, on average, equal to
TVaR0.95 [X] .
3 – Tail Value-at-Risk 26/41
Properties of TVaR
Tail Value-at-Risk is increasing in pI p < q⇒ TVaRp [X] ≤ TVaRq [X] .
Positive homogeneous and translation invariant:I For a > 0 and b ∈ R : TVaRp [aX + b] = aTVaRp [X] + b.
I Exercise: Prove these properties.
TVaR is a coherent risk measure.I We give a prove in the next chapter.
3 – Solvency 27/41
Solvency capital requirement:I X is a future loss/liability.
I V = technical provision
I K [X] = solvency capital
K [X] = TVaRp [X]−V.
Interpretation:I K [X] = TVaRp [X−V] ≥ VaRp [X−V] .
I X−V = loss which is not covered by the technical provisions
P[X−V > VaRp [X−V]
]≤ 1− p.
I First buffer = VaRp [X−V] .
I Second buffer = TVaRp [X−V] .
3 – Other risk measures 28/41
Conditional Tail Expectation:I For any p ∈ (0, 1) and any r.v. X, the Conditional Tail Expectation at
level p isCTEp [X] = E
[X | X > VaRp [X]
].
I VaRp [X] ≤ CTEp [X] .
Expected Shortfall:I For any p ∈ (0, 1) and any r.v. X, the Expected Shortfall at level p is
ESFp [X] = E[(
X− VaRp [X])+
].
I Solvency capital = Value-at-Risk + Expected Shortfall.
3 – Relation between VaR, TVaR, CTE and ESF 29/41
Theorem
For p ∈ (0, 1) , we have that
TVaRp [X] = F−1X (p) +
11− p
ESFp [X] ,
CTEp [X] = F−1X (p) +
1
1− FX
[F−1
X (p)]ESFp [X] ,
CTEp [X] = TVaRFX[F−1X (p)] [X] .
Remarks:
I In general, FX
[F−1
X (p)]
is not always equal to p.
I If FX is a continuous distribution function:
CTEp [X] = TVaRp [X] .
3 – The normal distribution 30/41
X is normal distributed with mean µX and variance σ2X.
Value-at-Risk:I For p ∈ (0, 1)
VaRp [X] = µX + σXΦ−1 (p) .
Stop-loss premium:
E[(X− K)+
]= σXφ
(K− µX
σX
)− (K− µX)
[1−Φ
(K− µX
σX
)].
Expected Shortfall:
ESFp [X] = E[(
X− VaRp [X])+
]= σXφ
(Φ−1 (p)
)− σXΦ−1 (p) (1− p) .
3 – The normal distribution 31/41
Conditional Tail Expectation:I For p ∈ (0, 1)
CTEp [X] = F−1X (p) +
1
1− FX
[F−1
X (p)]ESFp [X]
= µX + σXφ(Φ−1 (p)
)1− p
.
I TVaRp [X] = CTEp [X] .
3 – Exercise 32/41
Consider the r.v. X with distribution function
FX (x) =
x , if 0 ≤ x < 0.85,
0.85 , if 0.85 ≤ x < 0.9,0.95 , if 0.9 ≤ x < 0.95
x , if 0.95 ≤ x ≤ 1.
Value-at-Risk:I Prove that the value-at-risk is given by:
Varp [X] =
p , if 0 < p ≤ 0.85,0.9 , if 0.85 < p ≤ 0.95p , if 0.95 < p ≤ 1.
.
Conditional Tail Expectation:I Take p = 0.9. Prove that
CTEp [X] = 0.975.
4 – The required solvency capital 33/41
Regulators view:I V = technical provisions of a future loss X.
I Solvency capital: ρ [X]−V.
I ϕ : risk measure to calculate the shortfall risk.
I Cost of insolvency:ϕ[(X− ρ [X])+
]Investors view:I Holding an extra amount of capital requires a return equal to ε :
(ρ [X]−V) ε.
Cost function:I For ε ∈ (0, 1) :
C (X, ρ [X]) = ϕ[(X− ρ [X])+
]+ (ρ [X]−V) ε.
I Find the optimal capital requirement K such that the cost functionC (X, K) is minimal.
4 – Optimal capital requirement 34/41
Theorem
Take 0 < ε < 1. The cost function C (x, ρ [X]) defined by
C (X, ρ [X]) = E[(X− ρ [X])+
]+ (ρ [X]−V) ε,
reaches its minumum if
ρ [X] = VaR1−ε [X] ,
4 – TVaR as a minimum 35/41
Minimal value of the cost function:
C (X, VaR1−ε [X]) = ε (TVaR1−ε [X]−V) .
Take ε = 1− p. For any K :
C (X, K) ≥ (1− p)(
TVaR [X]p −V)
.
So we find:
TVaRp [X] ≤ K +1
1− pE[(X− K)+
].
Note that we have an equality if K = VaRp[X].
4 – TVaR as a minimum 36/41
Theorem
TVaR is a minimum:
TVaRp [X] = minK
K +
11− p
E[(X− K)+
].
5 – The Haezendonck-Goovaerts risk measure 37/41
Consider an insurer, facing the loss X ∈ (0, max[X]).
Total available capital (including a solvency buffer)= V.
Capital with extra risk buffer = ρ.
The random variable Z is defined as:
Z =(X−V)+
ρ−V.
I If X ≤ V, there is no extra capital needed and Z = 0.
I If Z ≤ 1, the extra buffer ρ−V absorbs the unanticipated losses.
I If Z > 1, the extra buffer ρ−V is not sufficient to cover realized loss.
5 – The Haezendonck-Goovaerts risk measure 38/41
We force E[Z] to be equal to 1− α, where α ∈ (0, 1).
The solution is then denoted by ρ[X, V]:
E
[(X−V)+
ρ[X, V]−V
]= 1− α.
We can then write:
ρ[X, V] = V +1
1− αE[(X−V)+
].
We determine the capital level V, such that ρ[X, V] is minimal.
5 – The Haezendonck-Goovaerts risk measure 39/41
The linear Haezendonck-Goovaerts risk measure ρ[X] is definedas:
ρ[X] = infV∈[0,max[X]]
ρ[X, V].
We can write:
ρ[X] = infV∈[0,max[X]]
V +
11− α
E[(X−V)+
]We find that:
ρ[X] = TVaRα[X].
6 – Risk measures and comonotonic sums 40/41
VaR, TVaR, ESF are additive for comonotonic risks
Theorem
Consider a comonotonic random vector (Xc1, Xc
2, . . . , Xcn) and let
Sc = ∑ni=1 wiXc
i , where wi are positive weigth factors. Then we have for allp ∈ (0, 1) that
VaRp [Sc] =n
∑i=1
wiVaRp [Xi] ,
TVaRp [Sc] =n
∑i=1
wiTVaRp [Xi] ,
ESFp [Sc] =n
∑i=1
wiESFp [Xi] .
Exercise: prove this theorem.
6 – CTE and comonotonic risks 41/41
Continuous distribution functions:I Assume that the cdf of Xc
i are continuous.
I Then we have that CTEp [Xi] = TVaRp [Xi] .
I Sc has also a continuous distribution function: CTEp [Sc] = TVaRp [Sc] .
I CTE is additive for comonotonic risks.
In general: CTE is not additive for comonotonic risksI counter example: FY is uniform over [0, 1].
I FX is
FX (x) =
x , if 0 ≤ x < 0.85,
0.85 , if 0.85 ≤ x < 0.9,0.95 , if 0.9 ≤ x < 0.95
x , if 0.95 ≤ x ≤ 1.
I For p = 0.9 : CTEp [Sc] < CTEp [X] +CTEp [Y] .