Chapter 2 Risk ordering 2.1 Position of the problem Since insurance is used to reduce the risk, we need an objective criteria in order to compare risk. Unfortunately, it is sometimes very difficult to compare risky situations : most of the time, agents have a different opinion. Even in a non-risky situation, in front of several goods, the comparison between two profile of consumption is difficult and based on the concept of preferences (or sometimes of utility function) that will be recalled in the next subsection. 2.1.1 Individual preferences Let us consider the case of a discrete risky model described by Ω = {σ1 , . . . , σS }, for which the consumption set of agent i, denoted by Xi is either1 RS+ (or RS++ which is by definition, (]0, +∞[)S ). Note that for each event s, the consumption a(s) is a real number. The most classical modelization of the tastes of a consumer is the concept of total order associated to the agent. When the preferences of agent i are total, in front of two alternatives profile of consumptions, a or b, he can either • be indifferent between a and b (denoted by a ∼i b) • prefer a to b (denoted by a #i b) • prefer b to a (denoted by a ≺i b) It is important to distinguish2 the “weak” preference relation %i from the “strict” preference relation ≺i . Concretely, a %i b means either that a %i b and b &%i a. “The agent wants a at least as much as b” or “the agent weakly prefers a to b” Conversely, if the primitive data is the weak preference relation, we can build the asymmetric part, The choice of RS+ is the most natural “allowing any non-negative consumption” while restricting to strictly positive consumptions avoids a lot of mathematical difficulties about the boundary of Xi since it is an open set. 2 In mathematics, the “weak” preference relation is a non-strict (or reflexive) partial order. In these contexts while the strict relation is irreflexive. 1 9 10 CHAPTER 2. RISK ORDERING • a ≺i b ⇔ a %i b and b &%i a. • a ∼i b ⇔ a (i b and a %i b. 2.1.2 Utility function Since preferences are a mathematical theoretical concept, it is most of the times replaced by a numerical criteria. We say the a numerical function from Xi to R is a utility function that represents the preferences %i of the agent i if a %i b ⇔ u(a) ≤ u(b). It is important to understand that the utility function is not unique Proposition 2.1 If u represents the preferences %i then for any ϕ increasing function from R to R, v defined by v = ϕ ◦ u also represents the preferences %i . ! 2 For example, if X = R++ , u(x) = 3 x21 x2 and v(x) = 2 ln(x1 ) + ln(x2 ) are associated to the same preferences. A classical example of utility function in a stochastic context is the class of Von-Neumann Morgenstern utility function, where the agent is in front of a risky situation where the probability of state σk is pk . In this case, there exists some function ui such that Ui (X) = Ui (X(σ1 ), . . . , X(σS )) = p1 u(X(σ1 )) + . . . + pS u(X(σS )) If X is not stochastic “X = (α, . . . , α)”, then Ui (X) = u(α). Von-Neumann Morgenstern utility function are expected utility function. Risk-aversion, risk-neutral, and risk-loving behaviour are related to the concavity property of u. 2.1.3 dominance Let us start by an example, event σ1 σ2 probability 1/2 1/2 a 1 41 b 14 15 As told previously, it is reasonable to consider that every agent will have an opinion about “Is a better then b?”, but this opinion is “agent-dependent”. Therefore, if we want a notion of stochastic dominance, it will be a partial order between random variables. There are several notions of stochastic ordering, indeed one has to make an arbitrage between being able to compare a lot of situations, and being “accepted” by a lot of people. 2.2 2.2.1 Several introductory notions of stochastic dominance State-wise stochastic dominance The simplest case of stochastic dominance is state-wise dominance (also known as state-by-state dominance), defined as follows: 2.2. SEVERAL INTRODUCTORY NOTIONS OF STOCHASTIC DOMINANCE 11 Random variable a is state-wise stochastic dominant over random variable b gives at least as good a result in every state (every possible set of outcomes), and a strictly better result in at least one state. For example, if a dollar is added to one or more prizes in a lottery, the new lottery state-wise dominates the old one because it yields a better payout regardless of the specific numbers realized by the lottery. Similarly, if a risk insurance policy has a lower premium and a better coverage than another policy, then with or without damage, the outcome is better. Anyone who prefers more to less (in the standard terminology, anyone who has monotonically increasing preferences) will always prefer a state-wise dominant gamble. This notion will be “accepted” by everybody but as a counterpart, it is unable to compare the situation studied in subsection 2.1.3: neither a (SSD b nor a %SSD b. Even worst, it is also unable to order a and c. event σ1 σ2 probability 1/2 1/2 a 1 41 c 52 2 Also note, that two random variables following the same law are not similar with respect to SSD ordering. 2.2.2 First stochastic dominance In order to encompass the last example, we can introduce the first dominance defined as follows: a (F SD b if and only if for any t, Fa (t) ≤ Fb (t). Here clearly, c #F SD a. Since it involves the cumulative distribution function, two random variables following the same law are similar with respect to FSD ordering. Unfortunately, we can not conclude for the next exemple event σ1 σ2 probability 1/2 1/2 a 0 30 b 15 15 More generally, it is easy to show that a %F SD b implies that E(a) < E(b), so it is unable to compare two lotteries sharing the same expectation. 2.2.3 Mean-variance dominance Let X and Y be two random variables, and let us assume that E(X) = E(Y ), and consider them to be claim to be paid by insurance companies to its clients. We may consider that X dominates Y if V (X) ≤ V (Y ). This notion is inadequate in this context since it may underestimate potential risk. Let us consider an insurance company owning R = 50 as reserve funds in front of the alternative. event σ1 σ2 σ3 probability 0.5 0.499 0.001 X 20 0 0 Y 9.9 9.9 109.9 12 CHAPTER 2. RISK ORDERING Clearly, Y is less risky than X with respect to mean-variance (and they have the same expectation), but in real life, it is a much more dangerous situation since with positive probability, the insurance company may go to bankruptcy. This is why it is necessary for an actuary to pay attention to extreme values (even if they are with small probabilities and “hidden” by the criteria of variance) in order to estimate the dangerousness. 2.3 2.3.1 SMPS dominance definition Definition 2.2 Risk Y is said to have thicker tails than risk X (or to be riskier than X with respect to the Simple Mean Preserving Spread property) if • E(X) = E(Y ) • there exists some t0 in R such that " P (X ≤ t) ≤ P (Y ≤ t) if t < t0 P (X > t) ≤ P (Y > t) if t ≥ t0 We will write that X (SM P S Y , X is less risky than Y for the SMPS criterium (Y has been obtained by a Simple Mean Preserving Spread3 ). This presentation is centered on the notion of tails but " most of the time, we will write the FX (t) ≤ FY (t) if t < t0 definition in terms of cumulative distribution functions: FX (t) ≥ FY (t) if t ≥ t0 Proposition 2.3 Let X and Y be two random variables, and let us assume that E(X) = E(Y ), X (SM P S Y if and only if " FX (t) ≤ FY (t) if t < t0 FX (t) ≥ FY (t) if t > t0 Proof 2.4 It suffices to use the right continuity at point t0 . Once again, it involves only the cumulative distribution function, two random variables following the same law are similar with respect to SMPS ordering and it is the only case of similarity: X #SM P S Y , X is strictly less risky than Y for the SMPS criterium if X (SM P S Y , and FX &= FY . 2.3.2 Discrete examples Let us first study a simple example, event σ1 σ2 probability 1/2 1/2 X 1 1 Y 0 2 3 This notion is not introduced here even if the vocabulary will be used. 13 2.3. SMPS DOMINANCE FY 1 FX 0 0 1 2 Figure 2.1: A first SMPS comparison For t0 = 1, we can check that X #SM P S Y (cf. Figure 2.1). FX (t) = " if t < 0 0 1/2 if t ∈ [0, 2[ FY (t) = 1 if t ≥ 2 0 if t < 1 1 if t ≥ 1 The preference for a sure quantity is a very general result as shown by the following exercise. Exercise 2.5 Let Y be a non constant random variable and X be constant, equal to E(Y ). Show that X #SM P S Y . Here there was a unique possible value for t0 but the next example will show that it may be non unique. event σ1 σ2 σ3 probability 1/3 1/3 1/3 X 1 2 3 Z 0 2 4 0 1/3 FX (t) = 2/3 1 if if if if 0 1/3 FY (t) = 2/3 1 t<1 t ∈ [1, 2[ t ∈ [2, 3[ t≥3 if if if if t<0 t ∈ [0, 2[ t ∈ [2, 4[ t≥4 Since for any t ∈ [2, 3[, FX (t) = FY (t), any value of t0 in [2, 3] could be used (cf Figure 2.2). 1 FX FY 0 0 1 2 t0 3 Figure 2.2: A second SMPS comparison 4 14 CHAPTER 2. RISK ORDERING Exercise 2.6 Let us consider X and Y . Show that SMPS. event σ1 σ2 probability 1/4 1/4 X 10 20 Y 9 20 X is more risky than Y with respect to σ3 σ4 1/4 1/4 30 40 30 41 Exercise 2.7 Let us consider X and Y . Show that X #SM P S Y (transfert of probability from center to the tails). event σ1! σ2! σ3! σ4! probability .3 .2 .2 .3 Y 10 20 30 40 event σ1 σ2 σ3 σ4 probability .2 .3 .3 .2 X 10 20 30 40 2.3.3 Density examples In the case of density variables, most of the times, the picture is easier to analyse (Figure 2.3) but once again, t0 may be non unique (Figure 2.4). 1 FY FX 0 0 1 2 Figure 2.3: A typical SMPS comparison, here X (SM P S Y 1 FY FX 0 0 1 2 Figure 2.4: A non typical SMPS comparison 2.3.4 Classical properties Proposition 2.8 X #SM P S Y if and only if −X #SM P S −Y . Exercise 2.9 X #SM P S Y if and only if −X #SM P S −Y . Proposition 2.10 Let a and b be real numbers such that a &= 0, then X #SM P S Y if and only if aX + b #SM P S aY + b. Proposition 2.11 Let ϕ be an increasing function from R to R, then X #SM P S Y if and only if ϕ(X) #SM P S ϕ(Y ). 15 2.3. SMPS DOMINANCE Lemma 2.12 Let X be a real valued random variable subset Γ of Ω with positive probability, let us defined Y and sharing the same expectation. Formally, " X(ω) Y (ω) = E(X1Z )/P (Z) defined on a set Ω, let us consider any as constant on Γ, equal to X outside Γ if ω ∈ /Γ if ω ∈ Γ If X is not constant on Z, then Y #SM P S X. Proof 2.13 Let us denote t0 the value of Y on Γ. Note that X = Y on both sets Y < t0 = {ω ∈ Ω | Y (ω) < t0 } and Y > t0 . • Let t < t0 , we have FY (t) = P (Y ≤ t) = P ((Y ≤ t) ∩ ΓC ) (since Y = t0 on Γ) = P ((X ≤ t) ∩ ΓC ) (since Y = X on ΓC ) ≤ P (X ≤ t) • Let t > t0 , we have FY (t) = 1 − P (Y > t) = 1 − P ((Y > t) ∩ ΓC ) (since Y = t0 on Γ) = 1 − P ((X > t) ∩ ΓC ) ≥ 1 − P (X > t) ≥ FX (t) (since Y = X on ΓC ) Lemma 2.14 Let X and Y be a real valued random variable defined on a set Ω, such that E(X) = E(Y ). Let us assume that there exists a real number t0 and a partition (Ω+ , Ω? , Ω− ) of Ω such that • For all ω ∈ Ω+ , t0 < X(t) ≤ Y (t); • For all ω ∈ Ω? , X(t) = t0 ; • For all ω ∈ Ω− , Y (t) ≤ X(t) < t0 . Then X %SM P S Y . Proof 2.15 Let us distinguish the two cases: • Let t < t0 , we have FX (t) = P (X ≤ t) = P ((X ≤ t) ∩ Ω− ) + P ((X ≤ t) ∩ Ω+ ) + P ((X ≤ t) ∩ Ω? ) = P ((X ≤ t) ∩ Ω− ) (since X > t on Ω? ∪ Ω+ ) = P ((Y ≤ t) ∩ Γ+ ) (since Y ≤ X on Ω− ) ≤ P (Y ≤ t) = FY (t). 16 CHAPTER 2. RISK ORDERING • Let t > t0 , we have FX (t) = 1 − P (X > t) = 1 − (P ((X > t) ∩ Ω− ) + P ((X > t) ∩ Ω+ ) + P ((X > t) ∩ Ω? )) = 1 − P ((X ≤ t) ∩ Ω+ ) (since X ≤ t on Ω? ∪ Ω− ) ≥ 1 − P ((Y ≤ t) ∩ Γ+ ) (since Y ≤ X on Ω+ ) ≤ P (Y ≤ t) = FY (t). Exercise 2.16 Deduce Lemma 2.12 from Lemma 2.14. 2.3.5 Characterization in terms of quantile functions Proposition 2.17 Let X and Y be such E(X) = E(Y ). Then X (SM P S Y is and only " that −1 FX (p) ≥ FY−1 (p) if p ≤ p0 ; if there exists some p0 ∈ ]0, 1[ such that FX−1 (p) ≤ FY−1 (p) if p > p0 . Once again the condition at point p0 could be omitted. If the cumulative distribution functions are one to one, the proposition can be understood graphically (Figure 2.5). 2 t0 p0 0 FY−1 FX−1 ! ! ! ! ! ! FX " ! ! " ! " FY ! ! ! 0 p0 t0 2 Figure 2.5: quantile characterization Lemma 2.18 Let X and Y be real valued random variable such that E(X), E(Y ), V (X) and V (Y ) are finite, then X %SM P S Y implies that V (X) ≤ V (Y ). Proof 2.19 It suffices to show that E(Y 2 ) ≥ E(X 2 ) or that E((Y − a)2 ) ≥ E((X − a)2 ) for some a since E(X) = E(Y ). From Proposition 2.17,there exists some p0 ∈ ]0, 1[ such that " −1 FX (p) ≥ FY−1 (p) if p ≤ p0 ; FX−1 (p) ≤ FY−1 (p) if p > p0 . We can consider a = FY−1 (p0 ), one gets both " −1 |FX (p) − a| ≥ |FY−1 (p) − a| if p ≤ p0 ; |FX−1 (p) − a| ≥ |FY−1 (p) − a| if p > p0 . Consequently, |FX−1(p) − a| ≥ |FY−1(p) − a| for any p. We can conclude by an application of Proposition 1.13. 17 2.3. SMPS DOMINANCE Exercise 2.20 Prove again Exercise 2.6 using quantile characterization. Exercise 2.21 Prove again Exercise 2.7 using quantile characterization. 2.3.6 Application to stop-loss insurance The setting will be the following: let us consider a decision maker which is either a client of an insurance company who wants to insure himself or an insurance company that wants to reinsure itself. The decision maker has an initial wealth w > 0 (real number) and faces a risk which leads to possible losses X ≥ 0 (stochastic) such that E(X) is finite. He can buy an insurance contract defined by its profile of reimbursement X ≥ I ≥ 0, priced using a shared loading factor (the premium π(I) = (1 + λ)E(I)). The kind of insurance contract, he is allowed to use, is limited to this class. Finally, we assume that the decision maker is averse to simple mean preserving spread: in front of two risky situations B and C, if B (SM P S C, the decision maker will prefer B. Proposition 2.22 Let I be some initial insurance contract, then there exists some parameter d such that the associated stop-loss insurance contract is preferred to his initial contract. Proof 2.23 An indemnity Id is a deductible insurance contract of parameter d if Id (ω) = (X(ω) − d)+ := max(X(ω) − d, 0). The sketch of the proof is • determine the value of d • check that the final wealth if he uses the contract Id is less risky with respect to SMPS than the initial final wealth for the parameter t0 = d. For the first step, d will be chosen in such a way that the price of the stop loss insurance will be the same as the initial one. Since 0 ≤ I ≤ X, one gets that 0 ≤ E(I) ≤ E(X). Let us consider an auxiliary function ϕ : R+ → R+ defined by ϕ(d) = E(Id ) = E(X − d)+ . It is easy to show that ϕ is continuous and non-increasing, and satisfies ϕ(0) = E(X) and limd→+∞ ϕ(d) = 0. We can apply intermediate value theorem to get the existence of some d (non necessarily unique) such that ϕ(d) = E(I). For such a value of d the price πd of the stop loss insurance will be the same as the initial one π. For the second step, one has to compare the final wealth Zd if he uses the contract Id is less risky with respect to SMPS than the initial final wealth Z. " Z(ω) = w − π − X(ω) + I(ω) Zd (ω) = w − πd − X(ω) + Id (ω) In view of Proposition 2.10, one has the equivalence Zd (SM P S Z ⇔ Yd (SM P S Y where Y = w − π − Z and Yd = w − π − Zd . These variables Yd = X − (X − d)+ and Y = X − I represents the amount of damage uncovered by the insurance contract. In particular " X(ω) if X(ω) ≤ d Yd (ω) = d if X(ω) > d 18 CHAPTER 2. RISK ORDERING Since the two insurance contracts share the same premium, they have the same expectation and therefore E(Y ) = E(Yd ). In order to conclude, let us discuss4 with respect to t • If t < d, then in one hand, the event Yd < t is a subset of the event X < d (case of small damage), and on this set Yd = X, consequently, FYd (t) = FX (t). On the other hand, since I ≥ 0, Y = X − I ≤ X and FY (t) ≤ FX (t) (in fact true, for any t). We can conclude that in this case, FYd (t) ≤ FY (t). • If t > d, then the event Yd > t is impossible. Therefore FYd (t) = 1 ≥ FY (t). 2.3.7 A missing property 1.0 1.0 FX 1.0 FZ 0.5 0.5 0.5 FZ FY FY 0 FX 0 0 2 0 0 2 0 2 Figure 2.6: FX and FY 1.0 1.0 FX 1.0 FZ 0.5 0.5 0.5 FY 0 0 0 2 0 0 2 0 2 Figure 2.7: FX and FY Let us analyze the left part of Figure 2.6, clearly since the sign FX −FY changes three times, it is impossible to compare X and Y with respect to SMPS order. But if we introduce a third variable Z(Figure 2.7), then we can understand that Z (SM P S Y (middle part of Figure 2.6) and X (SM P S Z (right part of Figure 2.6). This means that the concept of SMPS ordering is not transitive which shows that we need to introduce a new concept since as noticed by Rotschild and Stiglitz, a definition of “greater risk” should be transitive. 4 An alternative proof of this point can be done by applying Lemma 2.14. 19 2.4. MPS DOMINANCE 2.4 2.4.1 MPS dominance Lemmas Lemma 2.24 If E(X) = E(Y ), then the following properties are equivalent ' p −1 ' p −1 F (u) du F (u) du X i) for any p ∈ ]0, 1[, 0 ≥ 0 Y p p ' 1 −1 ' 1 −1 FY (p) dp FX (p) dp p p ≤ ii) for any p ∈ ]0, 1[, 1−p 1−p Proof 2.25 Let us first remark that the denominator can be omitted. Let us introduce two auxiliary functions: ( 'p 'p ϕ(p) = 0 FX−1 (u) du − 0 FY−1 (u) du '1 '1 for any p ∈ [0, 1], ψ(p) = p FX−1 (u) du − p FY−1 (u) du In view of our assumption, ϕ(p) + ψ(p) = E(X) − E(Y ) = 0. Consequently they have opposed signs. Lemma 2.26 If X (SM P S Y , then for any p ∈ ]0, 1[, ' p −1 ' p −1 F (p) dp F (p) dp X 0 ≥ 0 Y p p Proof 2.27 With the notations of Lemma 2.24, we want to show that ϕ is non negative. Since ϕ! (p) = FX−1 (p)−FY−1 (p), we already know that ϕ is non-decreasing on [0, p0 ] and non-increasing on [p0 , 1]. But ϕ(0) = ϕ(1) = 0, it suffices to fill the table in order to get the sign. p p0 0 ϕ! (p) + 1 − ϕ(p) 0 2.4.2 0 Definition Definition 2.28 The random variable X is less to be less risky than Y in the sense of Mean Preserving spread order (or of the stop loss order) if and only if ) p ) p −1 FX (u) du FY−1 (u) du E(X) = E(Y ) and for any p ∈ ]0, 1[, 0 ≥ 0 . p p 20 CHAPTER 2. RISK ORDERING It is easy to check that the ordering (M P S is transitive. Clearly, in view of Lemma 2.26, one has also X (SM P S Y ⇒ X (M P S Y. It is also easy to check that the ordering (M P S is transitive. This definition can be interpreted as a ranking about the worst case, which is contexte dependent. Let us consider the p = k/100. If X and Y are financial assets, it means that “the average payoff to be expected among the k% smallest possible payoff is greater with X than with Y ” while in the context of claims “the average claim to be to paid in case of claims belonging to the (100 − k)% greatest ones is smaller with X than with Y ”. An agent i will be said strongly strict averted if X (M P S Y ⇒ X (i Y. Let us now pay attention to the discrete case when the n outcomes are equiprobable . Let us assume that the possible values are ranked: x1 < x2 < . . . < xn and y1 < y2 < . . . < yn . Then X (M P S Y is equivalent to x1 + x2 + . . . + xn = y1 + y2 + . . . + yn x1 ≥ y1 x1 + x2 ≥ y1 + y2 ... ≥ ... x1 + x2 + . . . + xn−1 ≥ y1 + y2 + . . . + yn−1 Exercise 2.29 Let us consider this distribution, compare in the sense of MPS order X and Y . event σ1 σ2 σ3 σ4 probability 1/4 1/4 1/4 1/4 X 3 4 7 6 Y 8 2 4 3 2.4.3 Characterization Theorem 2.30 (Rotschild and Stiglitz, 1970) Let us assume that E(X) = E(Y ), then the following statement are equivalent: • X (M P S Y ; • there exists a sequence (X0 , . . . , Xn , . . .) of random variables such that X0 = X, for each n, Xn+1 (SM P S Xn and Xn →n→+∞ Y . (convergence in law5 ) As particular case, in the case of the example treated in Subsection 2.3.7, it is possible to build the following sequence: X0 = X, X1 = Z, and for each n ≥, Xn = Y . Since for each Xn+1 (SM P S Xn , we can conclude that X (SM P S Y . Proposition 2.31 (Rotschild and Stiglitz, 1970) Let us assume that E(X) = E(Y ), then the following statement are equivalent: 5 pointwise convergence of the characteristic functions, or equivalently convergence of FXn (t) to FY (t) when FY is continuous at point t. 2.4. MPS DOMINANCE 21 i) X (M P S Y ; 'x 'x ii) for any x, −∞ FX (t) dt ≤ −∞ FX (t) dt. iii) for any d, E((X − d)+ ) ≤ E((Y − d)+ ) If in addition, X an Y are assumed to be bounded then the following statement are equivalent: i) X (M P S Y ; iv) for any u non decreasing and concave, E(u(X)) ≥ E(u(Y )). v) for any u concave, E(u(X)) ≥ E(u(Y )). Remark 2.32 Condition ii) is known in the literature as Second Order Stochastic Dominance, so if E(X) = E(Y ), then X (SSD Y is equivalent to X (M P S Y . Clearly, X (F SD Y implies X (SSD Y . Remark 2.33 Condition iii) means that the premium of a stop-loss insurance will be lower. Exercise 2.34 Let us consider X following a normal law N (m, 1), Y = 2X and Z = −X. • Compare X and Y with respect to MPS order. • Compare X + Z and Y + Z with respect to MPS order.