Introduction ELEMENTS OF GENERAL RISK THEORY Vladimir Rykov (E-mail: vladimir rykov@mail.ru) Dept. of Appl. Math. and Comp. Modelling, Russian State University of Oil&Gas Introduction OUTLINE 1. Introduction 2. Background 2.1 2.2 2.3 2.4 2.5 Risk notion Risk measurement Some classes of r.v.’s and operations with them Some parametric families of risk distributions Hierarchical systems and event tree 3. Engineering Risk Theory 3.1 3.2 3.3 3.4 3.5 3.6 Methodology of risk analysis Structural reliability Risk tree construction Risk tree analysis Qualitative analysis 4. Insurance Risk Theory (Ruin Problem) Introduction Introduction During last years the term “RISK” became very popular in different areas: mathematics, I engineering, I economics, I environment, I management, I biology & medicine, etc. and different authors use this term in different senses. I Below we consider some definitions of risk in different sources. Introduction Dictionaries. I I [Webster’s new universal unabridged dictionary (1996)] “Risk is chance of losses, probability of losses” [Ushakov commenting dictionary (1935)] “Risk is possible dangerous, possible loss (damage) or non-success in business”. Introduction International documents. I I [Basic safety principles for nuclear power plants. 75-INSAG-3 (1999)] “Risk R(t) for time t, connected with some event, is defined as a product of the probability of this event by non-wished consequence D of this event: R(t) = P(t)D” [Functional safety of electrical/ekectronic/programmable electronic-related systems (ISO/IEC Guide 51: 1990)] “Risk is combination of the probability of occurrence of harm and the severity of that harm”. Introduction National laws. I I I [Russian Federation law about technical regulation (2002)] “Risk is a probability of harm causing for life or healthy of citizens, property of physical or juristical persons, states or municipal property, environmental, life or healthy of animals and plants with taking into account the size of this harm” [Russian Federation Federal Law No.7-FL.] “Ecological risk is a probability of events occurrence, that has some bad consequences for environment and caused by negative influence of economical or other activity, extraordinary situations of natural and technogeneous character.” [Law of Ukraine (2005)] “Risk is a possibility of occurrence and a probable scale of consequences from negative phenomenon during some time period”. Introduction Normative documents. I I I [Russian State Standard. Risk Management. Terms and Definitions (2005)] “Risk is a combination of probability of event and its consequences”. [Russian State Standard. Safety aspects. Rules for including in standards (2002)] “Risk is a combination of harm probability and its heaviness”. [An American National Standard (ANSI/IEEE, 1987). IEEE Guide for General Principles of Reliability Analysis of Nuclear Power Generation Station Safety Systems] “Risk is a measure of the probability and severity of undesired effects. Often taken as the simple product of probability and consequence”. Introduction Authors. I I [Alpeev A. (2005)] “Risk is an evaluation of expected damage from possible event in given conditions”. [Efimov S.L. (1996)] “Risk is 1) a danger of a negative event”...”... Risk is characterized by a collection of circumstances in their unity and interaction”, 2) an object of insurance”, 3) a kind of responsibility of the insurer”. Introduction I [Commercial insurance in gas industry. M.: “Gasprom” (1998)] “A wide spectrum of ... circumstances, not dependent from will of the owner, having a property of probability, and unexpectedness of appearance, are called as risk of economic activity of the subject”. Introduction I I [E.J. Henley, H.Kumamoto (1992)] use a dictionary definition of risk: “Risk is the possibility of loss or injury to people and property” [J. Grandel (1991)] In the book under title “Aspects of risk theory” there is no definition of risk at all, only descriptions of some risk models. Introduction I I [Gordon B.G. (2003).] “Risk R is the product of probability of event P by the value of its consequences V : R = PV ”. [Glenn Koller (2005)] Risk is “. . . A pertinent event for which there is a textual description. . . . Associated with each risk are typically at least two parameters: probability of occurrence, consequence (impact) of occurrence.” Introduction I [N. Singpurvala (2006)] Term risk arises in many contexts, but there is not its definition and methods of measurement. I [E. Solojentsev (2006)] Also term risk arises in many contexts, there are many parameters of risk but its definition is also absent. Introduction Thus, the situation in this area now looks like a situation in reliability theory more then half a century ago, when different specialists use the term “reliability” in different senses and understood under reliability different its characteristics. Introduction Historically among mathematicians and actuaries this term arisen in framework of ruin models, Cramer (1930, 1955), Grandel (1991), a.o. From another side, among engineers the notion risk is used in framework of reliability theory and means the losses or damages for peoples, environment and finance, Heynly&Kumamoto (1992). There is one more conception of risk due to von Niemann and Morgenstern (1953) which use utility theory for comparison of random variables (r.v.). It is used for comparison of different financial and business decisions utilities. Introduction Therefore, now there exist at least three different risk theories: I Insurance risk theory, which deals with the ruin problems; Engineering risk theory, which deals with investigation of the consequences (damages), generated with failures of complex non-reliable systems; I Business risk theory, which considers the problems of r.v. comparison in framework of utility theory. In this situation the strong and common for different cases notion of risk, that could be used in inter-disciplinary connections, and the methods for the risk measurement and assessment are needed. I Introduction The analysis of existing notion risk definitions shows that I any authors also as different national and international documents define this term quite differently, including into definition different characteristics of risk. I I This shows the necessity to propose the general definition of risk understandable and useful for different kind of specialists. Introduction I. BACKGROUND Chapter I. BACKGROUND I.1. Risk notion Different kinds of risks pursue both individuals during whole their life and juridical subjects: I industrial, I agricultural plants, I financial, I insurance companies, I societies, I political communities, I biological objects, I ecological structures etc. Introduction Most of economical, technical, political social and others decisions connected with “risk”, that is inevitably arise due to randomness and uncertainty of factors influencing to the phenomenon to which the decision is taken. Habitual idea about “risk” means an occurrence some random “risk events”, and its consequence such as financial, material or others loses. To separate a common component and differences in different situations, which in habitual life are joint as risks situations, to show the problems of general risk theory, and to formulate the goal and mean notions of general risk theory, consider several examples. Introduction 1.1. Examples 1.1.1. Risk of the health lost Crossing the street in the icy time one has a risk to fall and twist a leg or to get another injury. Here risk arise as a (random) event, and its consequence is the cost of treatment, or missing reward. 1.1.2. Risk of work ability lost Some types of professions contains the risks of lost of able to work before the retire time as a result of professional illness. Another types of work-ability lost risks arise as a result of technological or social changing in the community. It is obvious that the able-bodied lost also leads to the financial and moral loses. Introduction 1.1.3. Property missing risks Driving one has a risk of accidence. Opening new business one has a risk to loss the money, if it is not going well. Some person or a factory risk with its wealthy in the case of fire, storm, earthquake etc. These phenomenons are also distributed in time and also brings some loses. 1.1.4. Social risks Because there are something to lose for the owners, they need to spend some money for the life, health, property etc. defence. Nevertheless, the adequate social organization of community might be more direct and reliable way for defence of the interests of different layers of society. The problem what this organization should be is not the question for this course. Another example of social risks are medical risks. Introduction 1.1.5. Medical risks For surgical operation we are risking with some disability and even the life in more degree with less experienced surgeon. Nevertheless, as a Russian writer and physician V.Veresaev noted, that because the experience comes only with practice, thus if nobody will use the young surgeons, the whole community risks to lost the professional surgeons. This example shows a contradiction between an individual and social risks. Introduction 1.1.6. Financial and business risks The play at a stock-market also as any financial operation associated with risks. 1.1.7. Natural, ecological and technogeneous risks The nature itself from one side and the humans’ activity from the other side are sources of risks. The avoidances on large chemical enterprizes, breaks in oil- and gas- pipelines etc. represent significant hazard and lead to high damage for population and environment. Since these risks are connected with reliability of appropriate equipment, their study directed to excuse the payment for the providing and support of the necessary level of reliability equipment. Introduction 1.1.8. Insurance of risks and risk of insurance Using the idea of risks accumulation (collectivization) the insurance serves to stabilization of economics by means of smoothing of the risks consequences. From another side an insurance itself is a source of risks, which is connected with claims payment. More about insurance risks will be done later. The most of examples above deals with separate (individual) risks. But it is necessary to take into account that in many of the above examples (especially in insurance problems) risk situations can be repeated in time that leads to the necessity to study so called collective risks. Introduction The variety of these examples demonstrates the difficulties of general risk theory construction. To do that one should separate the common part of all considered (and many others) situations, to formulate the notion of risk and to propose the measure and the tools for measuring and comparison of risks. Introduction Linguistic analysis and definition of risk notions The linguistic analysis of the examples above shows the in any risk situation the expert deals with NEGATIVE CHANGES of CONDITION of an OBJECT under inner or exter nal This analysis justifies the following substantial definition AFFECTS or CIRCUMSTANCES Introduction Definition Definition Risk is an event or sequence of events, which leads to negative changes in an object states under inner or external affects or circumstances. These events occurs in time and accompany with different damages. This definition is not enough strong, but allows to propose some general mathematical approach to model any risk situation. Introduction 1.2. Notion of risk In all of these examples risk is connected with occurrence some I uncertain event A, which is called risk event I from the family F of events, describing considered risk situation, and it is characterized by two values: the time up to the event occurs (T ), and I the size of damage (X ). I Introduction There are different types of uncertainties in the world. Accordingly to Kolmogorov approach probabilistic type of uncertainty is characterized by two main restrictions: the possibility (at least in principle) infinitely many times observe the phenomenon considered, I under homogeneous conditions. In this case statistics is the real tool for probability estimation. I Introduction Therefore, from mathematical points of view the situation could be considered with different methods: I I I I probabilistic, subjective (by methods of subjective probabilities), expert (by utility theory methods), by the fuzzy sets technique, etc. Introduction In the following we limited ourselves mostly by probabilistic aspects and methods of risks study. From this point of view the risk should be described as a probabilistic space (Ω, F, P) on which a two-dimensional r.v. (T , X ), or a sequence of two-dimensional r.v.’s (Sn , Xn , n = 0, 1, . . . ) are determined. Definition A risk is a two-component r.v. (T , X ), or a sequence of two-components r.v.’s (a marked point process), (Sn , Xn , n = 0, 1, . . . ), defined under some probabilistic space (Ω, F, P). Remark. The second component X might be a multi-dimensional one or even functional. Therefore, the term random element would be more appropriate, but we will use the first more usual and simple. Introduction The given definition seems a very simple, however some peculiarities of risk analysis should be taken into account. These peculiarities are: I the uniqueness of any risk situation; the long chain of causes, that is necessary to be taken into account in risk situation analysis; I absence of needed statistical data. All these circumstances lead to the necessity to construct for any risk model its own probability space, which is usually realized in terms of the risk tree construction, that will be considered later. I Introduction 1.3. Classification of risk models The above examples give the possibility to classify risks accordingly to their generation. But this approach does not help for the general theory development, since it distracts from general notions and methods of risk analysis. Therefore in further we focus on general problems of risk modelling and analysis. The main problem of risk analysis consists in risk events uncertainties. There are many types of uncertainties in the world, and the problem of investigation and measurement of uncertainty is studied by many scientists (see Bibliographical notes). We restricted ourselves in this lectures mostly by the probabilistic one. Introduction There are two possible approach for random phenomenons modelling: I theoretical (analytical); I statistical. The first one is certainly good, but it has limited applications, because it is based on the delicate analysis of the phenomenon considered, that not always possible. The second one is theoretically universal, nevertheless not always practically applicable because of necessity collection and elaborating enough large information, which not always accessible because of its cost and many of risk events are very rare. Therefore in risk theory we need also in I subjective (on the base of subjunctive probabilities and expert analysis). This approach has an evident defect, that is its subjectivity. Introduction Risk models traditionally are divided into: I individual risks models and I collective risks models. An individual risk is risk, connected with one-point risk event though possibly from wide family of events, while A collective risk model deal with sequence of risks events, occurred in time jointly with its damages. These phenomenons are studied in the framework of risk processes. Introduction We will also divide individual risks into: I simple risks and I compound risks. A simple risk is characterized by the possibility to evaluate or to estimate its probabilistic characteristics directly. A compound risk is characterized with many different events and leads to numbers consequences. In this case it is natural to represent it as a chain of components (simple risks), each of which could be described in the framework of simple individual risks. Introduction In mathematical terminology I An individual risk is probabilistic model (Ω, F, P), on which two-component r.v. (T , X ) is defined, I I I the first component T represents the time up to risk event A occurrence beginning from some fixed epoch, the second one X denotes a damage from this risk event. A collective risk is probabilistic model (Ω, F, P), on which it is defined a sequence of two-component r.v.’s, {(Sn , Xn ) : n = 0, 1, . . . } I I the first components Sn represent the time up to n-th risk event An occurrence beginning from some fixed epoch, the second ones Xn denote damages from these risk events. Therefore construction and study of the probabilistic risk model for different risk situations compose one of aspects of the mathematical risk theory. Introduction The time T of risk event occurrence should be measured from some natural “initial” epoch t0 of the process beginning. In reliability theory this epoch is the equipment beginning exploitation epoch. I In the life insurance models appropriate epoch is the birth epoch. I Nevertheless, in many practical situations (for example in some catastrophic events in nature: earthquakes, tsunami, etc.) it is impossible to fix some specific beginning epochs. Note also that if a risk situation is studied in a fixed time interval, then the risk event can not occurs during this interval at all. I Introduction Therefore, three type of risk models should be considered. I I I Short-time model, where the probability of risk event occurrence is mach more smaller then one Middle-time model, where the probability of risk event occurrence is smaller then one. Long-time model, where the probability of risk event occurrence equals to one. Introduction Risks Classification RISKS CLASSIFICATION - NATURE METHODS engineering, environmental, production, financial, social, insurance, others. - stochastic, subjective, expert, fuzzy. TIME - short, - middle, - long. MODELS - individual, - compound, - collective. Introduction 1.4. Bibliographical notes. The problem of investigation and measurement of uncertainty is studied by many scientists beginning from Kardano (1501-1575), including D.Bernoulli (1700-1782), De Moivre, De Finetty, Fermat, Huygens, Laplace, Pascal, Poisson and others. Kolmogorov (1903-1987) specifies probabilistic uncertainty that is characterized by two maim conditions: I possibility to observe a phenomenon (in principle) infinity many times, I in homogeneous conditions. In this case the problem of uncertainty is described in terms of probability space (Ω, F, P) and mathematical statistics is a tool for the probability measure estimation. We will refer to this type of uncertainty to as randomness while safe the term uncertainty for all other its types. Introduction Historically the notion risk among of engineers arisen in framework of reliability theory and means the losses or damages for peoples, environment and finance. The methodology of engineering risk analysis firstly was proposed in Bell Laboratory (for bibliography see Henley and Kumamoto, 1992). Insurance risks are studied mostly in terms of ruin problem. The ruin problem also has a long history. The first works in this direction appears due to Lundberg (1903) and in the framework of mathematical insurance models belongs to H.Cramer, 1930. For some other details and approaches see also H.Cramer, 1955, S.Andersen, 1957, W.Feller, 1966, S.Asmusssen, 1996, Jan Grandell, 1991, and others. The conception of risks comparison in the framework of utility theory belongs to J. von Niemann and O. Morgenstern, 1953. The new results and the bibliography in this direction one can find in the review of V.Rotar and V.Bening, 1994. Introduction I.2. Risk measurement In this section the main risk measure and some special its characteristics will be proposed. Introduction 2.1. Risk distribution Above an individual risk was defined as a two-component r.v., both of which without of lost generality can be supposed to be positive. Therefore, it is determined by its cumulative distribution function (c.d.f.) F (t, x) = P{T ≤ t, X ≤ x}. (1) Remark. The damage itself is a multi-dimensional value, and moreover, for some applications it can be functional. Therefore, it can be considered as an element in some complex (may be functional) space. Nevertheless, further we will limited ourself mainly with positive r.v. for damage. Introduction In most practical situations the information about joint distribution of time and damage of risk event is not accessible, and we need to limited ourselves with only marginal c.d.f. of times FT (t) = P{T ≤ t} = F (t, ∞), (2) FX (x) = P{X ≤ x} = F (∞, x). (3) and damage size Later in § 5 some parametric families of distributions for times and damages description will be done. Introduction If risk is considered at fixed time interval, then instead of risk event occurrence time T it is more convenient to consider the probability P(A) of the risk event A and a conditional c.d.f. of damage given A, G (x; A) = P{X ≤ x|A} with G (0; A) = P{X = 0| A} = 0. In this case unconditional damage size has a form with a jump at zero, because there is no damages if risk event does not occurs, FX (x) = 1 − P(A) + P(A)G (x; A)). Introduction In general, it is more convenient to measure risk by the distribution of risk event time occurrence FT (t) = F (t) and conditional risk damage distribution given T G (x; t) = P{X ≤ x|T = t}. (4) Therefore, their joint distribution is Z t F (x, t) = G (x; u)dFT (u). (5) 0 In the simplest case it is supposed that time and damage are independent. G (x; u) = G (x) = FX (x) and F (x, t) = G (x)F (t). Introduction In many practical cases this assumption is quite admissible with only the remark that the future damages at given time should be discounted with some discount rate s and therefore the future damage is measured with its present value, given with X̂ = e −sT X . Really, to cover the damage of the size X after the time T it is enough to put in the bank sum X̂ under s%. Therefore the c.g.f of damage present value is Z ∞ F̂X (x) = P{X̂ ≤ x} = P{Xe −st ≤ x} dFT (t) = 0 Z ∞ st = FX (xe ) dFT (t). (6) 0 Introduction Everywhere later in this course it is supposed that r.v. T and X are independent, and we will denote their c.d.f. as F (.), and G (.) respectively. In reliability (and engineering risk) theory the tail of the c.d.f. FT (t) R(t) = 1 − FT (t) = P{T > t} (7) is usually called by reliability function (in demography and insurance it is called with survival function). Following to these traditions it will be called here by risk function. Introduction For continuous distributions these functions plot at the figure below 1 0.9 F (t) 0.8 0.7 0.6 0.5 0.4 0.3 0.2 R(t) 0.1 0 0 1 2 3 4 5 6 7 8 9 10 t Figure: C.d.f. time to risk event and reliability function. Introduction For continuously observed r.v. it is more convenient to model their distributions with the help of probability density function (p.d.f.) f (t) = F 0 (t) and g (x) = G 0 (x). (8) Therefore, the c.d.f. can be represented in terms of p.d.f. in the form Z x Z x F (x) = f (u) du and G (x) = g (u) du. 0 0 In most practical cases the time and the damage are measured in discrete units. In this case an appropriate model is discrete distribution fk = P{T = k}, gk = P{X = k} (9) where k represents the number of time or damage units. Discrete distributions are also used for modelling compound distributions in § 6. Introduction 2.2. Risk intensity and hazard rates Since the times to risks events are measured usually from some special time (or event), it is very important to know conditional distribution of the residual time to risk event occurrence, given time t after “natural” beginning of the process. This is the conditional probability that the risk event occurs in time interval (t, t + x], given that it does not take place before the time t. (In the formula below it should not be confused with two-dimensional distribution (1)). FT (x; t) P{t < T < t + x} = P{T > t} F (t + x) − F (t) R(t) − R(t + x) = . (10) 1 − F (t) R(t) = P{T ≤ t + x|T > t} = = Introduction For the small values of x one can get: F (x, t) = f (t) x = λ(t)x. 1 − F (t) More strongly the function λ(t) is defined by the formula 1 F (t + ∆) − F (t) f (t) = , ∆→0 ∆ 1 − F (t) 1 − F (t) λ(t) = lim (11) where function λ(t) represent conditional “instantaneous” probability density of risk events occurs at time t after the natural beginning of the process and it is called risk hazard rate function. The integral of it Z t f (u) Λ(t) = du = − ln(1 − F (t)) (12) 0 1 − F (u) is called a hazard function. Remark. In reliability theory the function λ(t) is known as hazard rate, and in demography it is known as mortality rate. Introduction With risk hazard rate function one can evaluate the probability of risk event occurrence during a small interval ∆t at the time t after its natural beginning time with the area under the curve as it is shown in the figure 2.2, where typical mortality rate function is presented. Enough high value of this function at the very beginning shows the enfant mortality level. λ(t) λ t t+∆ Figure: Hazard rate function. t Introduction The equality (12) allows to represent c.d.f. of time to risk event and the risk function in terms of risk hazard function Z t 1 − F (t) = R(t) = exp{− λ(u) du} = exp{−Λ(t)}. (13) 0 Analogously, for the probability of risk occurrence in the time interval (t, t + x] one can find t+x Z F (x; t) = P{T ≤ t + x| T > t} = exp − λ(u) du . (14) t Introduction Remark. Jointly with the h.r.f. in time in different models it might be useful to consider also the hazard rate in space. It might be called as intensity (rate) of risk expansion. For example it is possible to suppose that the forrest fire expansion has a downward risk rate expansion. This means that the probability of localization and stopping of fire at the beginning stage is enough high, and it decreases with its expansion up to some limit, defined by another factors, after that it could increase again. Introduction 2.3. Generating functions and their properties. The generating functions is a very useful tools for different characteristics of risk calculation. Definition Let G is a distribution of a non-negative r.v. X , then the function Z∞ g̃ (s) = Ee sX Z sx = e G (dx) = 0 ∞ e sx g (x)dx 0 is called the moment generating function (m.g.f.) of r.v. X , and/or its c.d.f. G (x). Introduction Note that up to the sign in power it coincides with the Laplace transform of the p.d.f. or Laplace-Stiltjes transform of the c.p.f. of r.v. X . Z∞ g̃L (s) = Ee −sX = Z e 0 −sx G (dx) = 0 ∞ e −sx g (x)dx = g̃ (−s). Introduction Definition Let {pk , k = 0, 1 . . . } is a distribution of an integer-valued r.v. N, then the function ∞ X z k pk p(z) = Ez N = k=0 is called the probability generating function (p.g.f.) of r.v. N, and/or its distribution {pk , k = 0, 1 . . . }. Introduction Remark 2.1. The coefficients of Taylor expansion of m.g.f. at point s = 0 give the moments of r.v. X , and the coefficients of Taylor expansion of p.g.f. at point z = 0 produce up to known multipliers the distribution of r.v. N. Remark 2.2. In spite of m.g.f. and p.g.f. are defined differently for different r.v.’s, there exist a close connection between them, and they are connected with one more transformation of r.v. distributions, namely with characteristic functions. However, we will not touch of these connections, since in further we will not use it. In the context for both of these functions we will the term “generation functions” (g.f.). Introduction The application of these function is based on their properties, that represented in the following theorems. Theorem [2.1.] A m.g.f. g̃ (s) of any non-negative r.v. X is unique determined in all region Re s ≤ 0 of a complex variable s. A p.g.f. p(z) of any integer-valued r.v. N is unique determined in all region |z| ≤ 1 of a complex variable z. The distributions of appropriate r.v.’s are uniquely reconstructed with their g.f.’s. Introduction Theorem [2.2] G.f. of the sum of independent r.v.’s equals to the product of g.f. of summands. In other words, g.f. of the convolution of distributions equals to the product of their g.f.’s. Proofs of these theorems contain in any textbook on probability theory, and it is omitted here. ¥ Remark. These functions also can be considered as a functions of real value. In this case they are connected with absolutely and completely monotone functions. Introduction Definition A function f (z) of real value z is called absolutely monotone at the segment [a, b], if it is infinitely many times differentiable inside of this segment and f (n) (z) ≥ 0 for a < s < b; a function f (s) of real value s is called completely monotone at the segment [0, ∞], if it is infinitely many times differentiable for s > 0 and (−1)n f (n)(s) ≥ 0 for s > 0. Theorem [2.3] P.g.f. is analytical inside 0 ≤ z ≤ 1 and all its derivatives are positive, while LST of any non-negative r.v. is analytical for s ≥ 0 and all its derivatives are sign-alternating. In other words p.g.f.’s are absolutely and LST of any non-negative r.v. are complete monotone functions. Proof can be find in Feller, 1966, and it is omitted here. ¥ Remark. As an exercise it is proposed to transform this theorem also for m.g.f. Introduction Applications of the absolutely and completely monotone functions are based on the following theorems. Theorem [2.4] 1. If f and g are complete monotone functions, then fg is also complete monotone function. 2. If f is a complete monotone function, and g is a positive function with a complete monotone derivative, then the function f (g ) is also a complete monotone one. Proof should be fulfilled as an exercise. ¥ Introduction One more useful property of g.f.’s concerns to the calculation of sums of random number of r.v.’s. It is based on the following theorem which is a corollary from theorems 2 and 4. Theorem [2.5.] If p(z) is a p.g.f. of a integer-valued r.v N, and g̃ (s) is a m.g.f. of i.i.d. r.v’s Xi i = 1, 2, . . . independent of N, then the m.g.f. of compound r.v. N X Xi Y = i=1 equals to g̃Y (s) = p(g̃ (s)). (15) Introduction Proof is obtained by the simple calculation with the help of the complete probability formula, g̃Y (s) = Me −sY = ∞ X k=0 ∞ X £ pk M e −sY |N = k] = pk g̃ k (s) = p(g̃ (s)). ¥ k=0 These properties are widely used for different characteristics of compound distributions calculation, especially for calculation of the compound distributions moments. Introduction 2.4. Moments and other risk characteristics. Not only functional characteristics of risk, but also some numerical, especially mean value and variance for the time up to risk event, and appropriate damages are interesting in practice. Z∞ µT = ET = Z∞ Z∞ t f (t) dt = (1 − F (t) dt = R(t) dt 0 σT2 = 0 0 Z∞ DT = E(T − µT )2 = (t − µT )2 f (t) dt. (16) 0 Z∞ µX = EX = Z∞ x g (t) dx = (1 − G (x)) dx 0 σX2 = DX = E(X − µX )2 = 0 Z∞ (x − µX )2 g (x) dx. 0 (17) Introduction Remind that mean characterizes the “center of probabilistic weight” of a r.v., while the variance characterizes of dispersion of appropriate r.v. around its mean value. Remark 2.1. Analogous expressions for discrete r.v. should be presented in terms of appropriate sums. Remark 2.2. In finance mathematics the variance is often used as risk notion. The applications of these methods for the risk moments calculation will be considered in § 6. Introduction 2.5. Additions Examples. 1. Let {pk } be Poisson distribution with mean λ > 0, pk = λk −λ e , k = 0, 1, 2, ... k! Acc. to def. its p.d.f. is p(z) = e −λ(1−z) . Introduction 2. Let {pk } be Poisson distribution with mean λ > 0. Calculate the convolution g = p (1) ∗ ... ∗ p (n) Taking into account that {p (j) } equals to pj (z) = e −λj (1−z) and using the second part of the Th. 2 for g (z) one can find g (z) = n Y j=1 or pj (z) = k Y e −λj (1−z) , j=1 g (z) = e −(λ1 +...+λk )(1−z) . Now from the first part of the Theorem it follows that g is the Poisson distribution with mean λ1 + ... + λk . Introduction 3. Using generating functions method find compound damage. If ϕ(s) is a m.g.f. of damages Yi , then from part 2) of the theorem it follows that Z ∞ ∞ X X pk [ϕ(s)]k . pk e −sx g ∗k (x) dx = ψ(s) = k=0 k=0 Let p(z) = ∞ X pk z k k=0 be the p.g.f. of the damages number N. Then ψ(s) = p(ϕ(s)). Introduction Exercises. 1. Proof the first part of the Theorem 2.1. 2. Proof the Theorem 2.2. 3. Proof the Theorem 2.3. 4. Proof the Theorem 2.4. 5. Proof the Theorem 2.5. Introduction I.3. Some classes of r.v.’s and operations with them In this section some special operations with non-negative r.v.’s, and some classes of distributions such r.v.’s will be considered. Introduction 3.1. Some operations with r.v. and their distributions For risk analysis, especially for damage calculating some special operations under r.v.’s and their distributions are needed. In this section we consider some special techniques for this. 3.1.1. Shift. Some times jointly with r.v. X it is necessary to consider also r.v. Y = X + a, where a is some constant. Such an operation is called as a shift, the c.d.f. of shifted r.v. equals FY (x) = P{X + a ≤ x} = FX (x − a), and plotted at the picture below Introduction FHxL Λ=5, a=0.4 1 0.8 0.6 FX HxL=1-e-Λ x FY HxL=1-e-Λ Hx-aL 0.4 0.2 x 0.25 0.5 0.75 1 1.25 1.5 1.75 Figure: A shifted c.d.f. Introduction Example If somebody want to buy a house with help of some middleman with fixed payment a for help, his real price for the house will be distributed as a shifted by a price the house, Y = X + a. Introduction 3.1.2. Scaling. For the unit of measuring changing and/or for reducing different r.v. to the same unit of measure the operation of scaling is used. This operation consists in multiplying of a r.v. by some constant a. The c.d.f. of scaled r.v Y = aX equals ³x ´ FY (x) = P{aX ≤ x} = FX . a and it is plotted at the picture below Introduction FHxL Λ=5, a1 =0.4, a2 =1.4 1 0.8 FX HxL=1-e-Λ x x x FX H L=1-e-Λ a1 a1 x x FX H L=1-e-Λ a2 a2 0.6 0.4 0.2 x 0.2 0.4 0.6 0.8 Figure: A scaled c.d.f. Introduction Example If some damages X should be recalculated from euros to USD it should be multiplied by the current cost of euros in USD, for example a = 1.5. Therefore Y = 1.5X . Another situation of this type arise when a middleman want to take for his work a percents from the prise of the house from the previous example. In this case the real price of the house is a scaled by 1 + 0.01a initial price of the house. Introduction 3.1.3. Truncated distribution. To truncated distribution one refers to as a conditional distribution of a r.v. under condition that it takes its values only in some given subset of its previous values. This operation is used both in different theoretical calculation and in some applications. A distribution F (.) truncated, for example, to the segment [ab], is F[ab] (x) = P{X ≤ x|a ≤ x ≤ b} = FX (x) − F (a) 1{a≤x≤b} . FX (b) − FX (a) Introduction An appropriate graph at the figure below is shown FHxL 1 Λ=5, a=0.4, b=1.4 0.8 0.6 FX HxL=1-e-Λ x 0.4 e-Λ a - e-Λ x F@a,bD HxL= e-Λ a - e-Λ b 0.2 x 0.2 0.4 0.6 0.8 1 1.2 1.4 Figure: A truncated c.d.f. Introduction Example In some insurance agreements the damage X that is smaller than some fixed constant a does not covered. This constant is known as a “franchise”. In this case insurance claim Y equals to the damage truncated from the below Y = max{X , a}. Another type of this situation arise, for example, in reinsurance, when the main insurer take for himself risks with the damage not grater than a, and a reinsurer cover the claims which are grater than a. Then the main insure claim is the truncated from the above by a damage, and the reinsurer’s claim is truncated from the below by a damage. Their c.d.f. at the picture ? are shown Introduction 3.1.6. Mixture. At last consider the situation, when phenomenon of “risk” describes as some family of risk events, from which only one is realized in concrete situation. This is usual situation in engineering risk models, where several different risk events can occurs. Suppose that this family contains finite or denumerable set of disjoint events (any denumerable family can be reduced to disjoint one), F = {Ak : k = 1, 2, . . . }, each of which leads to its own damage Xk with c.d.f. Gk (x). Introduction Denoting with 1{Ak } the indicator functions of these events and with pk = P(Ak ) their probabilities, the resulting damage Y can be represented as ∞ X Y = Xk 1Ak . k=1 The c.d.f. G (x) of such damage is G (x) = P{Y ≤ x} = ∞ X k=1 and it is called by mixture of c.d.f.’s Gk (.). pk Gk (x) (18) Introduction Example An insurance company propose different covering in different accidents of life insurance: I illness, I dentist’s accidents, surgical operations, die, etc. If covering in these cases are r.v. Xk , then the full covering is their mixture with distribution I I G (x) = P{X ≤ x} = n X P(Ak )P{Xk ≤ x}. k=1 An example of mixture of two exponential distributions at the figure below is presented Introduction FHxL Λ1 =1, Λ2 =3, p1 =0.3, p2 =0.7 1 0.8 0.6 G1 HxL=1-e-Λ1 x G2 HxL=1-e-Λ2 x GHxL=p1 G1 HxL+p2 G2 HxL 0.4 0.2 x 1 2 3 4 Figure: A mixture of two exponential c.d.f. Introduction 3.1.7. Integrated tail distribution. The distribution 1 F (x) = µF Zx s (1 − F (u)) du, 0 R∞ where µF = 0 (1 − F (u)) du is the mean of distribution F is known as a integrated tail distribution for given distribution F . It is clear, that this distribution exists only for distributions with finite mean values. This distribution arise, for example, in reliability theory for stationary distributions of the “age” and the “residual life time” of elements in repairable systems, which is modelled as renewal processes. Remark. For exponential distribution the integrated tail distribution is the same. Introduction Example Consider accelerated reliability trails of some item with accelerated life time T distributed with p.d.f. F (t). Being failed each item instantaneously replaced by the new one. Then at some fixed enough long after the very beginning time t0 the residual time up to the fail has an integrated tail distribution Below the graph of an integrated tail distribution is done. Introduction FHxL Α=0.3, Β=3 1 0.8 x Α FHxL=1-e-I Β M 0.6 Α G H Α1 L - G I Α1 , I xΒ M M Fs HxL= G H Α1 L 0.4 0.2 x 200 400 600 800 1000 Figure: An integrated tail c.d.f. Introduction 3.1.4. Summation. If X1 , ..., Xk are independent r.v.’s with c.d.f.’s Gi (x) and p.d.f.’s gi (x), then the c.d.f. G (∗k) (x) of their sum Y = X1 + ... + Xk is given for all i = 1, k with the recursive formula Zx G (∗0) (x) = 1{x≥0} , G (∗i) G ∗(i−1) (x − u)gi (u)du, (x) = 0 which is known as a convolution of Gi (x) and is denoted by G (∗k) (x) = G1 ∗ G2 ∗ · · · ∗ Gk (x). (19) Introduction In the case of i.i.d. r.v. the last formula takes more simple form Zx G (∗0) (x) = 1{x≥0} , G (∗i) G ∗(i−1) (x − u)g (u)du. (x) = (20) 0 For discrete i.i.d. r.v. Xi with the common distribution gj = P{Xi = j} j = 0, 1, . . . the formula (20) is changed to the following one (∗0) gj = 1{j=0} , (∗k) gj = j X i=0 (∗(k−1)) gj−i gi . (21) Introduction Example If some insurance company has n agreements for a fixed period of time, and the claim for i-th agreement is Xi . Then the full claim of the company during this period equals to Y = X1 + · · · + Xn . Introduction 3.1.5. Compound distributions. The sum Y of random number N of random components Xi (i = 1, N) is known as compound r.v. and its distributions as compound distribution. Put ( 0, for N=0, Y = X1 + X2 + ... + Xk , for N=k=1,2,3.... Denote the distributions of r.v.’ N and Xk k = 1, N by pk = P{N = k}, k = 0, 1, 2, ... and Gk (x) = P{Xk ≤ x}. Introduction Then taking into account that the c.d.f. of the sum X1 + ... + Xk of fixed number of independent r.v.’s is G (∗k) (x) one can get P{Y ≤ x} = p0 1{x≥0} + ∞ X pk G (∗k) (x). (22) k=1 Notice that the index k here is not any more a r.v. but an integer, and the c.d.f. G (∗k) (x) is determined with the convolution of c.d.f. of r.v. X1 , ..., Xk . Introduction Example If some tourists firm organize a bus tour, and a risk event results with random number N claims, each of which is a r.v. Xk , k = 1, N. Then the whole payment of the insurance company is a sum Y = X1 + · · · + XN . We turn to this situation later, when compound damage distributions will be studied. For calculation of distributions of sum and compound damages it is necessary to use the convolution formula, which is not enough convenient for calculation. In these cases a convenient tools are characteristic or generating functions. These functions are used for modelling, analysis, and different characteristic calculation both time to risk event occurrence and values of damages. Introduction 3.4. Additions Exercises. Propose your own examples for I shifted, I scaling, I truncated, I mixture distributions Introduction II.4. Some parametric families of risk distributions One of the main problems of the risk theory is the modelling of I times to risk event arising and I damages from the event, which is based on the characterization properties of distributions. The nature of arising and development of risk situations is usually enough complex. Thus, it is expedient that jointly with simple “unite” risks it is also need to consider models of compound risks for enough complex situations. Because the same parametric families of distributions can be used for modelling of both time to risk event arising and the damage from it we consider some parametric families and classes of non-negative r.v. distributions, that can be used for both of them. Some special characterization properties of these distributions also will be considered. We start with some discrete distributions. Introduction 4.1. Some discrete parametric families of distributions The most of observations are fixed in discrete of time or damage units, therefore continuous distributions are only useful models to describe real situations. Thus, in applications for describing both the time up to risk event and the value of damage the different types of discrete distributions are used. Moreover, discrete distributions are used in models of compound damages for number of risks description. In table 4.1 at the end of this section some discrete distributions jointly with their main characteristics, which are often used in risk practice, are given. Consider some of them jointly with their g.f., means and variances. Introduction 4.1.1. Degenerated distribution The degenerated distribution is used for description of a non random r.v. Its c.d.f. has a stepwise graph with a jump of unit size at point b, G (x) 6 G (x) = 1{x≥b} (23) 1 0 b x Fig. 2.4.1.1. Degenerated distribution The m.g.f. mean value and variance for this distribution are g (z) = Ez X = z b , EX = b, DX = 0. Introduction Mixture of such types distributions allows to construct two-point distribution and any discrete distribution as well, that often used for risk characteristics modelling. The degenerated distribution and its mixture can be used for claims modelling in different insurance models. Example 4.1.1 Consider some contract of a short-time life insurance, that provides two risk cases: I a natural death – event A1 , and I the death resulted by some accident – event A2 . Suppose that the probabilities of these events and the claims provided by insurance company are: P(Ai ) = pi , bi , i = 1, 2. Then the claim p.d.f. will be the two point step-wise function, shown at the figure 2.4.1.1.1b. Introduction G (x) 6 1 p1 0 b1 b2 - x Fig.2.4.1.1b. Mixture of degenerated distribution Introduction 4.1.2. Uniform distribution The uniform distribution is characterized with fixed number, say, n of values of r.v., each of which has the same probability, P{X = i} = 1 n 1, n. (24) The m.g.f., the mean value and the variance of this distribution are Ez X = z 1 − zn , n 1−z EX = 1 (n + 1), 2 DX ≈ 4(n3 − 1) − 3(n + 1)2 . 12 The graph of c.d.f. has a step-wise form that is presented below at the figure 2.4.1.2. Introduction 1 F (x ) 5/6 2/3 1/2 1/3 1/6 x 0 -1 0 1 2 3 4 5 6 7 Figure: Fig.2.4.1.2. The graph of uniform distribution. 8 Introduction The uniform distribution can be used for damage distribution modelling in the cases, when there is not enough information. Example (4.1.2) Suppose that the repair of a car after an accident costed bi = $100i, (i = 1, 100), and all these payments are equally-probable. This damage is described with the uniform distribution. Introduction 4.1.3. Bernoulli distribution This distribution is used for description of two-valued r.v. By choose the scale and shift it is possible to reduce these values to zero and one, pk = p k (1 − p)1−k , k = 0, 1. Its g.f. equals p(z) = 1 − p(1 − z), and mean and variance are µ = EX = p, σ 2 = DX = p(1 − p). The sequence of i.i.d. Bernoulli distributed r.v.’s constitute so called Bernoulli scheme. The graph of this distribution is presented at the figure 2.4.1.3. (25) Introduction G (x) 6 1 1−p - x 0 1 Fig.2.4.1.3. Bernoulli distribution Introduction Example (4.1.3) Numerous observations during long time over all world show that for among any 1000 birth approximately 515 are boys and 485 are girls. Therefore, if one introduce a r.v., that takes value 1 for boy birth, and value 0 for girl birth, it will have a Bernoulli distribution with p = 0.515. Remark. Using shift and scaling operation any two point distribution also could be modelled with Bernoulli distribution. If B is Bernoulli r.v., then Y = b1 + (b2 − b1 )B, where is two point distributed r.v. see fig. 2.4.1.1b. b1 < b2 Introduction G (x) 6 1 p1 0 b1 b2 - x Fig.2.4.1.1b. Mixture of degenerated distribution Introduction 4.1.4. Geometric distribution This distribution is defined with the formula pk = (1 − p)p k , Its p.g.f. is p(z) = k = 0, 1, 2, . . . (26) 1−p 1 − pz and two first moments are µ = EX = p , 1−p σ 2 = DX = p . (1 − p)2 This distribution describes the moment of the “first success” in Bernoulli scheme. The graph of this distribution at the figure 2.4.1.4 is shown. Introduction g k (2/3) 0.3 0.2 0.1 k 0 5 10 15 20 Figure: Geometric distribution (p = 23 ). 25 30 Introduction Among discrete distributions this is the unique which possesses of the luck of memory property. Lemma (4.1.1) The luck of memory property P{T ≥ k + l|T > l} = P{T ≥ k} = p k (27) is a characterization property of geometric distribution among discrete distributions. Prove this lemma as an exercise. ¥ Introduction Example (4.1.4(a)) If some couple plan to have just one boy in their family, the number of children will be distributed geometrically with parameter p = 0.515. Example (4.1.4(b)) If in some telecommunication system a signal is transmitted with probability of error equals to q = 1 − p, them the length of right transmitted symbols will be geometrically distributed with parameter p Introduction 4.1.5. Binomial distribution The binomial distribution µ ¶ n k pk = p (1 − p)n−k , k k = 0, 1, . . . , n, depends on two parameters: natural n and real p. Its m.g.f. is p(z) = (1 − p(1 − z))n , and two first moments are µ = EX = np, σ 2 = DX = np(1 − p). This is the distribution of number of “successes” in Bernoulli scheme with n trials. (28) Introduction The graph of the distribution at the figure 5.1.5 it is presented. 0.16 b k (2/3) 0.12 0.08 0.04 k 0 0 5 10 15 20 25 30 Figure: The shape for binomial distribution (n = 30, p = 23 ). Introduction Example (4.1.5.) Some previous geological investigations show the presence of oil in given region with probability p. If it is planed to bore n oil holes, the number of non-empty oil holes among them will be described by binomial distribution with parameters n, p. Introduction 4.1.6. Negative binomial distribution The negative binomial distribution µ ¶ α+k −1 k pk = p (1 − p)α , k k = 0, 1, . . . (29) for integer α represents the distribution of the trails number k in Bernoulli scheme up to α-th success, but it also can be spread out of any real α. Its m.g.f. is ³ 1 − p ´α p(z) = 1 − pz and two first moments are µ = MX = αp , 1−p σ 2 = DX = αp , (1 − p)2 The graph of this distribution at the figure 2.4.1.6. Introduction 0.12 0.1 0.08 0.06 0.04 0.02 5 10 15 20 Figure: The Negative binomial distribution graph. Introduction Example (4.1.6.) If an Oil-drilling company want to investigate some region up to α-th non-empty hole, the whole number of holes needed satisfies to the negative binomial distribution with parameters α, p, where p is the oil presence probability in the considered region. Introduction 4.1.7. Poisson distribution This distribution is defined with the formula pk = λk −λ e , k! k = 0, 1, 2, . . . , where real λ > 0 is its parameter. Its m.g.f. is p(z) = E −λ(1−z) , and two first moments are µ = MX = λ, σ 2 = DX = λ. The following figure represents Poisson distribution. (30) Introduction 0.12 π k (15) 0.1 0.08 0.06 0.04 0.02 k 0 0 5 10 15 20 25 Figure: The Poisson distribution graph (λ = 15). 30 Introduction The main property are stability against summing, which is represented by the following lemma. Lemma (4.1.2.) The sum of two (and several) Poisson-distributed r.v.’s has the Poisson distribution parameter of which equals to the sum of parameters of summands. Prove this lemma as an exercise. ¥ Due to this stability property, concerning in previous lemma, the Poisson distribution is used in many practical cases. Introduction The above property admits the following generalization. Theorem (4.1.1.) The sum of “many small” i.i.d. discrete r.v. has approximately Poisson distribution. Formally, if P{Ni = 1} = Then lim P n→∞ X 1≤i≤n λ , n P{Ni > 1} = Ni = k = λk −λ e , k! 1 . n2 k = 0, 1, 2, . . . . Introduction Example (4.1.7.) Due to the stability property, contained in Lemma 2, the Poisson distribution is very good apt for describing different kinds of event flows: I number of particles in radio-nuclear disintegration, I I number of calls to telephone station, number of E-mails, arisen to some server, etc. In table 5.1 the summary of the discrete distributions jointly with its main characteristics is given. Table 5.1. Some discrete distributions and their characteristics Title Distribution Mean value Variance Poisson λk −λ e , k = 0, 1, . . . k! λ λ λ>0 Geomrtric k (1 − p)p , k = 0, 1, . . . p 1−p p (1 − p)2 0≤p≤1 Binomial µ ¶ n k p (1−p)n−k , k = 0, 1, . . . n, np k np(1 − p) Table 5.1. Prolongation Title Negative binomial Distribution µ ¶ α+k −1 k p (1 − p)α , k Mean value Variance αp (1−p)2 αp 1−p k = 0, 1, . . . , 0 < p < 1, α > 0 p(p+ln(1−p)) − (1−p) 2 ln2 (1−p) Logarithmic − pk , k = 0, 1, . . . k ln(1 − p) 0<p<1 −p × (1 − p) 1 ln(1 − p) Introduction 4.2. Some parametric families of continuous distributions Consider some standard parametric families of continuous non-negative distributions, which is often used for the time to risk events and simple damages modelling. In the tables 5.2, 5.3 at the end of this section some standard families of such distributions are shown. Bellow some comments for them are given. Introduction 4.2.1. Uniform distribution The uniform distribution is determined with its p.d.f. g (x) = 1 1{a≤x≤b} , b−a 0 < a < b, (31) which has a rectangular form (see fig. 2.4.2.1), that gives it another name, rectangular distribution. The constants a and b are parameters of distribution. Its c.p.d. is for t < a, 0, t−a F (t) = b−a , for a ≤ t ≤ b, 1, for t > b, Introduction The graph of this distribution is shown below. 1 F (x ) 0.8 0.6 0.4 0.2 p (x ) x 0 -10 0 10 20 30 Figure: C.d.f. and p.d.f. for uniform distribution. 40 Introduction The hazard rate function is for t < a, 0, 1 λ(t) = b−t , for a ≤ t ≤ b, does not defined, for t > b, Mean value and variance for this distribution respectively are MX = a+b , 2 DX = (b − a)2 . 12 Characterization property of this distribution is equali-probable of any values of r.v. at any equal subintervals inside of segment [a, b]. Introduction This distribution can be used both for description of the times of risk events occurrence and the damage value in situations, when only the possible boundary values for time or/and value of damage interval are known, and more detailed information about their distribution absents. I As examples it can be used as property insurance model, when only the boundaries for estimated property are known, but real value of claim unknown. I Another example could be the failure time of new equipment, when there is no enough statistical data for it estimation. Example (4.2.1.) Under absence of more detailed information and statistical data it is possible to suppose that the pipeline lifetime uniformly distributed on the interval [0,30] (years). Introduction 4.2.2. Exponential distribution This distribution has a c.d.f. F (t) = 1 − e −λt , (32) where λ is a parameter. Reliability (survival) function is R(t) = e −λt , (33) The mean value and the variance equal µT = MT = λ−1 , σT2 = DT = λ−2 . The graph of exponential distribution is presented at the figure 2.4.2.2. Introduction 1 0.8 F (x ) 0.6 0.4 p (x ) 0.2 x 0 0 2 4 6 Figure: Exponential distribution. 8 Introduction The hazard rate function is constant and equals to the parameter of distribution λ, f (t) λ(t) = = λ. (34) R(t) This property of the h.r.f. allows to consider this distribution as a model of instantaneous failures. Moreover this property is a characteristic property of the distribution Lemma (4.2.1.) Among continuous distributions the exponential distribution is only one with constant h.r.f. Proof. Really, the relation (34) shows that the h.r.f. is constant. The inverse follows from the formula Zt λ(u)du = 1 − e −λt . F (t) = − 0 ¥ Introduction Another characterization property of exponential distribution is its “luck of memory” property, which contains in the following lemma. Lemma (4.2.2.) An exponential distribution is only one for which the following “luck of memory” property holds P{T ≤ t + x|T > t} = P{T ≤ x} = 1 − e −λx (35) Introduction Proof. Really, using the formula for conditional probabilities one get P{T ≤ t + x|T > x} = = P{T ≤ t + x, T > t} P{t < T ≤ t + x} = = P{T > t} P{T > t} e −λt − e −λ(t+x) = 1 − e −λx = P{T ≤ x}. e −λt The inverse statement is obtained from R(t + x) = = P{T > t + x} = P{T > t + x|T > x}P{T > x} = P{T > t}P{T > x} = R(t)R(x). ¥ Introduction Exponential distribution can be used as for description of the time to risk events arise, as a model for damage. Due to its characterization properties in reliability theory it is used for description of completely random failures. Example (4.2.2.) Two girls talk by phone. After one subject they began to discuss another one and so on. In this case at any point of time the longevity of residual talk does not depend on the time spent to it up to now. Therefore, the whole longevity of talk can be described by exponential distribution. Introduction 4.2.3. Shifted exponential distribution The p.d.f. of this distribution is (see fig. 2.4.2.3) g (x) = λe −λ(x−b) 1{x≥b} , b ≥ 0. (36) Its mean value and variance are MX = b + 1 , λ DX = 1 . λ2 This distribution can be used both for the time till risk events and for the damages modelling, when appropriate values do not take values less than some given constant. For example in insurance it is used for claims modelling in the case when the claims less than given value do not compensated (franchise). Introduction 0.1 0.08 0.06 0.04 0.02 –10 0 10 20 30 40 50 x Figure: The shifted exponential distribution. 60 Introduction Example (4.2.3.) Automobile traffic at some not very loaded road can be described with Poisson flow, that means that the intervals between cars are exponentially distributed. Nevertheless, the movement rules demand to have some special intervals between cars. Therefore, the real distribution of the intervals between cars are shifted exponential one. Introduction 4.2.4. Gamma-distribution The p.d.f. of this distribution is (see fig. 2.4.2.4.) g (x) = λα x α−1 −λx e 1{x≥0} , Γ(α) α > 0, (37) R∞ where Γ(α) = 0 x α−1 e −x dx is the Gamma-function, and λ is a scale, while α is a shape parameters. For α = 1 it coincide with p.d.f. of exponential distribution. Survival (risk) function equals to Z∞ R(t) = F̄ (t) = λt x α−1 −x e dx Γ(α) Introduction 0.1 0.08 0.06 0.04 0.02 0 10 20 30 40 50 x Figure: P.d.f. of Gamma-distribution. 60 Introduction Appropriate c.d.f. is presented at the figure 17 1 0.8 0.6 0.4 0.2 0 10 20 30 40 x Figure: Gamma distribution. 50 60 Introduction The closed formula for the hazard rate function does not exists, but mean value and variance are µT = MT = α , λ σT2 = DT = α . λ2 The Gamma distribution has the following stability property. Lemma (4.2.3.) If αi , λ, then PTi : i = 1, 2 have the Gamma distributions with parameters P i Xi has the Gamma distributions with parameters i αi , λ. When α → ∞ Gamma distribution is approximated with normal one. Example (4.2.4.) The Gamma distribution is used for modelling non-symmetric unimodal distributions. Introduction 4.2.5. Lognormal distribution The survival function for this distribution is µ ¶ µ ¶ ln t − µ µ − ln t =Φ R(t) = 1 − Φ σ σ with p.d.f. g (x) = 1 √ xσ 2π e− (log x−µ)2 2σ 2 1{x≥0} , and mean and variance µX = MX = e µ+ σ2 2 , ¡ 2 ¢ 2 σX2 = DX = e σ − 1 e 2µ+σ . (38) Introduction Appropriate c.d.f. is presented at the figure 18 1 0.8 0.6 0.4 0.2 0 1 2 3 4 x Figure: Lognormal distribution. 5 6 Introduction 0.6 0.5 0.4 0.3 0.2 0.1 0 1 2 3 4 5 x Figure: p.d.f. of lognormal distribution. 6 Introduction Example (4.2.5.) In reliability theory this distribution is used for repair time modelling. From the other side, this is one of so called heavy-tailed distributions, and it is used for modelling “large” (catastrophic) damages of risk events. Introduction 4.2.6. Pareto distribution This distribution has a p.d.f g (x) = α ³ c ´α+1 1{x≥c} c x c >0 with expectation and variance µX = σX2 = αc for α > 1, α−1 αc 2 DX = for α > 2. (α − 1)(α − 2) MX = (39) Introduction 0.8 0.6 0.4 0.2 0 2 4 6 8 x Figure: p.d.f. of Pareto distribution. 10 Introduction Example (5.2.6.) Pareto distribution is also belongs to the class of heavy-tailed distributions, and it is used for large harms of risk event modelling. Introduction 4.2.7. Gnedenko-Weibull distribution The survival function for this distribution is α R(t) = e −λt , (40) with mean value and the variance µT = MT = σT2 = DT = ¡ ¢ Γ 1 + α1 1 λα , ¡ ¢ ¡ ¢ Γ 1 + α2 − Γ2 1 + α1 . 2 λα The exponential distribution being a special case of this distribution when α = 1. If T is a GW distributed r.v., then the r.v. X = λT α has an exponential distribution. Appropriate c.d.f is presented at the figure below Introduction 1 0.8 0.6 0.4 0.2 0 1 2 3 4 5 x Figure: The c.d.f. for Gnedenko-Weibull distribution. Introduction Its h.r.f. is λ(t) = αλt α−1 and it is presented at the figure. Figure: The h.r.f. for Gnedenko-Weibull distribution. Introduction The popularity of this distribution in Reliability Theory is explained by the property of this distribution to be as a limiting distribution for maximum and minimum if series of i.i.d. r.v.’s. Really, because the failure time of a series system is the minimum of failure times of its components, and the failure time of a system in parallel is the maximum of failure times of its components, the following theorem is very important. Denote by X(n) = max{Xi : i = 1, n}, and by X(1) = min{Xi : i = 1, n}. Introduction Because X(n) increases when n increases, it is natural to find some sequences of constants an , bn such that the limiting distribution of r.v.’s Yn = an X(n) + bn to be proper distribution. when n→∞ Introduction There are two types of limiting distributions for Yn dependently on the behavior of tails distribution of r.v.’s Xi , 1 − F (x) = P{X > x}. Let 1 − G (x) = lim P{Yn > x}. n→∞ Theorem (4.2.1.) The following holds: x I If 1 − F (x) ≈ e −x , when x → ∞, then 1 − G (x) = e −e . α I If 1 − F (x) ≈ x −α , when x → ∞, then 1 − G (x) = e −x . Introduction Proof. of this theorem is outside of the framework of this course, and it is omitted. ¥ The second case of this theorem leads to the GW distribution, while the first one corresponds to the Gompertz distribution, which in survival analysis is used. Introduction The GW distribution possesses the following stability property. Theorem (4.2.2.) If independent r.v. Xi , (i = 1, n) have the GW distribution with parameters (λi , α). Then the r.v. Y = min Xi 1≤i≤n P has GW distribution with parameters ( 1≤i≤n , α). Proof. can be got by direct calculation. ¥ Introduction Example (4.2.7.) The life time of a system in series with independent identical elements, life time of each of which has a GW distribution, also has a GW distribution α RS (t) = R1 (t) · · · Rn (t) = exp {− (λt α + . . . λt α )} = e −nλt . Introduction 4.2.8. Truncated normal distribution The truncated normal distribution can be used for description of the epochs for partially forecasting risks, especially in reliability theory for time to gradual failures. To explain this settings note that I if an item failure results of some physical parameter a goes out of limiting value a1 , and I I I this parameter changes accordingly to deterministic rule a = f (t, a0 ), from an initial random state a0 normally distributed, then the time up to the parameter a reaches to critical value a1 also has a normal distribution. Introduction Really, under described conditions the failure time is a solution of f (T , a0 ) = a1 . Denoting by ϕ(a1 , a0 ) the inverse function to f (t, a0 ), one get T = ϕ(a1 , a0 ). Expanding the function ϕ(a1 , a0 ) in the Tailor series with respect to a0 around the point a = Ma0 up to second order members, one get T = ϕ(a1 , a) + ϕ0a (a1 , a)(a − a0 ). From here it follows that if parameter a0 has a normal distribution, then the failure epoch T also must be normally distributed. However, because the time to failure has only non negative values, the distribution should be truncated at zero. Introduction The risk function for this distribution is ¡ ¢ ¡ ¢ 1 − Φ t−µ Φ µ−t σ ¢ σ¢ ¡ ¡ R(t) = = , 1 − Φ − σµ Φ σµ where Φ(x) = √1 2π Rx t ≥ 0, (41) u2 e − 2 du is the standard normal c.p.d. When (as −∞ usually holds) σ << µ, thus it is possible to use an approximation µ ¶ µ−t F̄ (t) = Φ . σ Expressions for the mean value and the variance of this distribution are enough complicated, but for the usual situation, when σ << µ they coincide with the parameters of normal distribution: µT = MT = µ σT2 = DT = σ 2 . Introduction The c.d.f. of the distribution is presented at the figure below 1 0.8 0.6 0.4 0.2 0 0.5 1 1.5 2 2.5 x Figure: The c.d.f. for truncated normal distribution. Introduction The h.r.f. is (t−µ)2 2 1 e − 2σ λ(t) = √ × ¡ µ−t ¢ , Φ σ 2πσ and as it can be shown numerically has an asymptote y = looks like as shown at the figure 2.4.2.8. (t−µ) σ and Figure: The hazard rate function for truncated normal law. Due to the h.r.f. property this distribution for degradation modelling is used. Introduction 4.3. Almost-luck-of memory distributions Some generalization of exponential family is so called family of Almost Luck of Memory (ALM) distributions. This family is used for modelling periodic phenomenons in life and nature. Some examples are: (i) periodicity in sun activity; (ii) periodicity of seasons in year, of ocean floods-ebbs, etc.; (iii) many risk events, such as auto accidents, fires, zunamis, epidemics, pollution, etc. Introduction We start with the defining the concept of lack-of-memory (LM) property for random variables. Definition A non-degenerated at zero non-negative r.v. X possesses the LM-property at point c > 0 iff P{X ≥ c + x | X ≥ c} = P{X ≥ x} for all x ≥ 0. Naturally, it make sense only if 0 < P{X ≥ c} < 1. Introduction Note also if this property holds for all possible values c of r.v. X , then accordingly to the lemmas 4.1.1 and 4.2.1 this r.v. has or geometric, or exponential distributions. Lemma (4.3.1.) If r.v. X possesses LM-property at the point c > 0, then this property holds for all points of sequence {am = mc}∞ m=0 . ¥ This means that the time axis can be divided into intervals [0, c), [c, 2c), . . . , [mc, (m + 1)c), . . . , in which the r.v. is renewed. Introduction The characterization property of the ALM distributions is Theorem (4.3.1.) Let r.v. X possesses of ALM property for sequence {cm = mc}∞ m=0 . Then (i) for continuous X its p.d.f fX (x), x ≥ 0 is fX (x) = (1 − a)a[x/c] fY (x − [x/c]c), (42) where a = P{X ≥ c} and fY (.) is p.d.f. of some continuous r.v. Y at finite interval [0, c); (ii) for discrete r.v.X its distribution pX (x) is pX (x) = (1 − a)a[x/c] pY (x − [x/c]c), (43) where the value a is the same as before, and pY (.) is distribution of some discrete r.v. Y with finite number of values {0, 1, . . . , c − 1}. ¥ Introduction Another description of the ALM family convenient for simulation is Theorem (4.3.2.) A r.v. X possesses the ALM property at point c and, therefore, for the sequence {cm = mc}∞ m=0 if and only if it can be represented as a sum X = Yc + cZ (44) of independent r.v.’s Yc and Z , such that Yc is distributed at finite segment [0, c), and Z has a geometric distribution pZ (z) = (1 − a)az , z = 0, 1, . . . . ¥ Most of these distributions can be used for modelling both the time till to simple risk events and the damage them resulted. Some parametric families of distributions jointly with their characteristics are represented at the tables 5.2, 5.3 in the end of this section. Introduction 4.4. Distributions with monotone hazard rate Some families of distributions are studied based on the properties of their hazard rate functions. I I For example constant h.r.f. characterized exponential or geometric distributions. This allow to model time to risk event arising with appropriate distributions if there exists some theoretical bases to suppose that their h.r.f. is constant. In other cases from some theoretical reasons or based on statistical observations one can suppose some special shape (form) of h.r.f. For example the long-term observations of the peoples’ mortality show some concrete shape of the h.r.f. for this phenomenon, which is shown at the figure 4.3.1. Introduction λ(t) λ t t+∆ Figure: Mortality rate function. t Introduction There exists an understandable treatment of this mortality rate behavior. Analogous behavior of h.r.f. is observed in another phenomena, for example for aging units in reliability theory, where the h.r.f. for aging units increases with its age. In most risk situations the risk event (or risk process) is a result of many different reasons. For these cases the models of compound distributions are more applicable for time till risk evens as well as for caused them damage. This class of distributions in the next section will be considered. Table 5.2. Reliability functions and their characteristics Title Reliability function failure rate hazard rate mean Variance value, Exponential e −λt λ λe −λt 1 , λ 1 λ2 µ, σ2 Normal ¡ ¢ Φ µ−t σ ¡ ¢ Φ σµ WeibullGnedenko e −λt α − √ (t−µ)2 2σ 2 − (t−µ)2 2σ 2 e e ¢ ¡µ¢ √ ¡ 2πσΦ σ 2πσΦ µ−t σ αλt α−1 e −λt α αλt α−1 ¡ ¢ Γ 1 + α1 , 1 λα ¡ ¢ ¡ ¢ Γ 1 + α2 − Γ2 1 + α1 1 λα Table 5.2. Prolongation Title Reliability function failure rate hazard rate mean value, Variance Gamma Z ∞ λt x α−1 −x (λt)α−1 −λt λ(λt)α−1 e −λt α , R∞ e dxλ e Γ(α) Γ(α) x α−1 e −λx dx λ λt α λ2 Lognormal µ Φ µ − ln t σ ¶ (ln t−µ)2 2σ 2 e− √ 2πσt Reley e Uniform − t2 2σ 2 t − t 22 e 2σ σ2 √ e− (ln t−µ)2 2σ 2 2π σt Φ 1 σ2 ³ µ−ln t σ e (µ+ ´ r σ2 2 π σ 2 ) Table 5.3. Damage distributions and their characteristics Title Uniform Shifted exponential P.d.f. Mean value 1 1{a≤x≤b} , b−a 0≤a<b λe −λ(x−b) 1{x≥b} , b > 0, λ > 0 a+b 2 1 b+ λ Gamma α α−1 λ x e −λx 1{x≥0} , Γ(α) α > 0, λ > 0 α λ Variance (b − a)2 12 1 λ2 α λ2 Table 5.3. Prolongation Title P.d.f. Mean value Variance Lognormal √ (ln x−µ)2 1 e − 2σ2 1{x≥0} , 2πσx e µ+ σ2 2 ¡ ¢ 2 2 e σ − 1 e 2µ+σ σ>0 Pareto α ³ c ´α+1 1{x≥c} , c x αc , α−1 c > 0, α > 0 α>1 Γ(p + q) p−1 x (1 − x)q−1 × Γ(p) Γ(q) p p+q αc 2 , (α − 1)(α − 2) α>2 Beta pq (p + q)2 (p + q + 1) Introduction 4.5. Additions Exercises. 1. Find c.d.f., p.d.f. the mean value and the variance for an insurance agreement, presented in the example 1. 2. Let the damage value has an exponential distribution with parameter λ and a p.d.f. g (x) = λe −λx 1x≥0 , λ > 0, Find the m.g.f., the mean value and the variance of that damage. Introduction Bibliographical notes Most of the material of this section in any textbook on Probability Theory can be found. About ALM-distributions see Chukova, Dimitrov, (1992), Chukova, Dimitrov, and Khalil, (1993), Dimitrov, Khalil (1992) Introduction II.5. Event Tree In this section the general concepts of hierarchical fault tolerance reliability systems and the event tree as a tools for their investigation. Introduction 5.1. Hierarchical Systems Before turning to Engineering Risk Theory, several words about general Hierarchical Systems and their analysis tools. Many of present-day complex technical systems and biological objects are characterized by the following main properties: I I hierarchical structure; implemented system of the states control. Hierarchy of structure means that the system consists of subsystems each of which is also divided on sub-subsystems etc. up to the lowest non-divided (elementary) level (see fig.1). We will refer to non-divisible part of system to as units. Usually hierarchy of structure leads to the property that the primary failures (faults) arise mainly at the lowest (elementary) level of the system, and gradually developing leads to the failure of blocks and subsystems of more higher levels, containing those elements. Introduction ® © System ­³P ª ³ PP ® ©³ ® © r r r S1,1 S1,n1 ­©H ª ­©H ª HH HH ©© ©© r r r r r r r r r r Um 1 r r Um k Um K Subsystems of level 1 Subsystems of any level Units Fig. 1. A complex multi-level hierarchical system. Introduction 5.2. Event Tree Failures of complex systems lead to essential damage, and their exploitation connects with risks. Event Tree construction and analysis is an appropriate tools for hierarchical systems analysis such as degradation and/or fault development. Event Tree is a tree-type graph, which root is a resulting event, and branches connects generating events Firstly it have been used in Bell Lab by H.A. Watson in the beginning of 60-th last century for system reliability analysis. But this approach could be used also in many others directions (risk analysis, medicine, etc.) Introduction In this course it’ll be used for Engineering Risk Analysis (ERA), and this means that the failures of appropriate systems will be considered as events in Event Tree. This mean that the event are failures of elements, subsystems and the whole system and the Event Tree will be interpret as a Fault Tree. For event tree construction due to Henley & Kumamoto (1992) some special symbols for event and gates are used. They are presented in the following Tables. Introduction Table: Event Symbols No. Event symbol Name Description 1 Circle Basic event with sufficient data 2 Diamond Undeveloped event 3 Rectangle Event represented by gate 4 Oval Conditional event used with inhibit gate 5 House House event. Either occuring or not 6 Triangle Treansfer symbols 7 Triangle Treansfer symbols Introduction Table: Gate Symbols No. Gate Symbol Gate Name Causal Relation Function 1 AND Output event occurs if all input events occur simultaneously f = 2 OR Output event occurs if any one of the input events occurs f =1− 3 Inhibit Input produces output when conditional event occurs f = XU Q Xi Q (1 − Xi ) Introduction Table: Gate Symbols. Prolongation 4 Priority AND Output event occurs if all input events occur in the order from left to right ? 5 Exclusive Output event occurs OR if one, but not both, of the input events occur f = X1 (1 − X2 )+ (1 − X1 )X2 m out of n f m 6 n Output event occurs if m out of n input = αi P Qn αi ≥m i=1 1−αi Introduction This general approach for the Engineering Risk Analysis is used and more detailed in the next Chapter will be considered.