Comparison of aleatory and epistemic uncertainty modelling, using @RISK Presentation at Palisade 2012 Risk Conference London, April 18-19, 2012 Hans Schjær-Jacobsen Professor, Director RD&I Copenhagen University College of Engineering Ballerup, Denmark +45 4480 5030 hsj@ihk.dk www.ihk.dk Agenda 1. 2. • • • 3. 4. 5. Alternative representations of uncertainty Four dialogues on uncertainty Rectangular and triangular representations Non-monotonic functions Correlation Representation and calculation Net present value case Conclusions Palisade 2012 Risk Conference Copenhagen University College of Engineering 2 Rectangular representation [a; b] and {µ; σ} Membership function Probability distribution Alternative interpretations h 1 α h = 1/(b-a) μ = (a+b)/2 σ2 = (b-a)2/12 α-cut 0 a Palisade 2012 Risk Conference b Copenhagen University College of Engineering 0 3 Trapezoidal representation [a; c; d; b] and {µ; σ} Membership function Probability distribution Alternative interpretations h 1 α h = 2/(b-a+d-c) μ = h[(b3-d3)/(b-d)-(c3-a3)/(c-a)]/6 σ2 = [3(r+2s+t)4+6(r2+t2)(r+2s+t)2 -(r2-t2)2]/[12(r+2s+t)]2 α-cut r 0 a Palisade 2012 Risk Conference s c t d Copenhagen University College of Engineering b 0 4 Triangular representation [a; c; b] and {µ; σ} Membership function Probability distribution Alternative interpretations h 1 α h = 2/(b-a) μ = (a+b+c)/3 σ2 = (a2+b2+c2-ab-ac-bc)/18 α-cut 0 a Palisade 2012 Risk Conference c Copenhagen University College of Engineering b 0 5 A Dialogue on Uncertainty (1) Model Owner (MO): How should I represent the independent uncertain variable x in my model? Uncertainty Specialist (US): What do you know about x? MO: I know that x can attain any value between a and b (a < b). US: Do you know more about x? MO: Not really. US: Then I suggest that x is best represented by a rectangular possibility distribution (an interval [a; b]). MO: What is a possibility distribution? I am used to work with probability distributions. US: But you do’nt know the probability distribution of x. MO: No, but in that case I think it is reasonable to assume that x is best represented by a rectangular probability distribution [a; b]. Palisade 2012 Risk Conference Copenhagen University College of Engineering 6 Rectangular and triangular arguments of x(1-x) 2.4 1.2 Probability Possibility 2.0 1.0 1.6 0.8 Probability 1.2 0.6 0.8 0.4 0.4 0.2 0.0 0.0 -0.2 0.0 Palisade 2012 Risk Conference 0.2 0.4 0.6 Copenhagen University College of Engineering 0.8 1.0 1.2 7 Function x(1-x) with rectangular argument x = [0; 1] 35 1.2 Probability Possibility 30 1.0 25 0.8 20 0.6 15 0.4 10 0.2 5 0 -0.05 0.00 0.05 0.10 0.15 0.20 0.0 0.30 0.25 x(1-x) Palisade 2012 Risk Conference Copenhagen University College of Engineering 8 A Dialogue on Uncertainty (2) Model Owner (MO): How should I represent the independent uncertain variable x in my model? Uncertainty Specialist (US): What do you know about x? MO: I know that the nominal value of x is c. Due to uncertainty x can attain any value between a and b (a < c < b). US: Do you know more about x? MO: Not really. US: Then I suggest that x is best represented by a triangular possibility distribution [a; c; b]. MO: What is a possibility distribution? I am used to work with probability distributions. US: But you do’nt know the probability distribution of x. MO: No, but in that case I think it is reasonable to assume that x is best represented by a triangular probability distribution [a; c; b]. Palisade 2012 Risk Conference Copenhagen University College of Engineering 9 Function x(1-x) with triangular argument x = [0; 0,2; 1] 35 1.2 Probability Possibility 30 1.0 25 0.8 20 0.6 15 0.4 10 0.2 5 0 -0.05 0.00 Palisade 2012 Risk Conference 0.05 0.10 0.15 Copenhagen University College of Engineering 0.20 0.25 x(1-x) 0.0 0.30 10 A Dialogue on Uncertainty (3) Model Owner (MO): How should I represent the independent uncertain variables x1 and x2 in my model? Uncertainty Specialist (US): What do you know about the variables? MO: They can attain any value between a1 and b1 (a2 and b2). US: Do you know more? MO: Not really. US: Then I suggest that they are best represented by rectangular possibility distributions. MO: I am used to work with probability distributions. US: But you do’nt know the probability distributions. MO: No, but in that case I think that the variables are best represented by rectangular probability distributions. US: Do you know about the correlation between the variables? MO: No, I think I will run a series of correlation coefficients. Palisade 2012 Risk Conference Copenhagen University College of Engineering 11 Alternative interpretations of a rectangular distribution 1.2 Probability 0.125 μ = (a+b)/2 σ2 = (b-a)2/12 Possibility 1.0 0.100 0.8 0.075 0.6 h = 1/(b-a) 0.050 0.4 0.025 0.2 0.000 0.0 5 a=7 Palisade 2012 Risk Conference b = 16 Copenhagen University College of Engineering 18 12 Addition of two rectangular distributions, correlation in per cent 0.16 1.2 Probability Possibility 1.0 0.12 0.8 0% 0.08 0.6 +100% 0.4 0.04 0.2 0.00 0.0 10 Palisade 2012 Risk Conference 15 20 25 Copenhagen University College of Engineering 30 Sum 35 13 Subtraction of two rectangular distributions, correlation in per cent 0.16 1.2 Possibility Probability 1.0 0.12 0.8 0% 0.08 0.6 -100% 0.4 0.04 0.2 0.00 0.0 -10 Palisade 2012 Risk Conference -5 0 Copenhagen University College of Engineering 5 Difference 10 14 Multiplication of two rectangular distributions, correlation in per cent 0.014 1.2 Probability Possibility 0.012 1.0 0.010 0.8 0% 0.008 0.6 +100% 0.006 0.4 0.004 0.2 0.002 0.000 0.0 0 Palisade 2012 Risk Conference 50 100 150 200 Copenhagen University College of Engineering 250 Product 300 15 Division of two rectangular distributions, correlation in per cent 1.8 1.2 Possibility Probability 1.6 1.0 1.4 0% 1.2 0.8 -100% 1.0 0.6 0.8 0.6 0.4 0.4 0.2 0.2 0.0 0.0 0.0 Palisade 2012 Risk Conference 0.5 1.0 1.5 Copenhagen University College of Engineering 2.0 Quotient 2.5 16 A Dialogue on Uncertainty (4) Model Owner (MO): How should I represent the independent uncertain variables x1 and x2 in my model? Uncertainty Specialist (US): What do you know about the variables? MO: I know that the nominal values are c1 and c2. Due to uncertainty any value between a1 and b1 (a2 and < b2) may be attained. US: Do you know more? MO: Not really. US: Then I suggest that they are best represented by triangular possibility distributions. MO: I am used to work with probability distributions. US: But you do’nt know the probability distributions. MO: No, but in that case I think that the variables are best represented by triangular probability distributions. US: Do you know about the correlation between the variables? MO: No, I think I will run a series of correlation coefficients. Palisade 2012 Risk Conference Copenhagen University College of Engineering 17 Alternative interpretations of a triangular distribution 1.2 0.25 Probability μ = (a+b+c)/3 σ2 = (a2+b2+c2-ab-ac-bc)/18 Possibility 1.0 0.20 0.8 0.15 0.6 0.10 α-cut 0.4 0.05 0.2 h = 2/(b-a) 0.00 0.0 5 a=7 Palisade 2012 Risk Conference c = 10 Copenhagen University College of Engineering b = 16 18 18 Addition of two triangular distributions, correlation in per cent 0.16 1.2 Probability Possibility 1.0 0% 0.12 0.8 +100% 0.08 0.6 0.4 0.04 0.2 0.00 0.0 10 Palisade 2012 Risk Conference 15 20 25 Copenhagen University College of Engineering 30 Sum 35 19 Subtraction of two triangular distributions, correlation in per cent 0.16 1.2 Probability Possibility 1.0 0% 0.12 0.8 -100% 0.08 0.6 0.4 0.04 0.2 0.00 0.0 -10 Palisade 2012 Risk Conference -5 0 Copenhagen University College of Engineering 5 Difference 10 20 Multiplication of two triangular distributions, correlation in per cent 0.014 1.2 Probability Possibility 0.012 0% 1.0 0.010 0.8 +100% 0.008 0.6 0.006 0.4 0.004 0.2 0.002 0.000 0.0 0 Palisade 2012 Risk Conference 50 100 150 200 Copenhagen University College of Engineering 250 Product 300 21 Division of two triangular distributions, correlation in per cent 1.8 1.2 Possibility Probability 1.6 1.0 0% 1.4 1.2 0.8 -100% 1.0 0.6 0.8 0.6 0.4 0.4 0.2 0.2 0.0 0.0 0.0 Palisade 2012 Risk Conference 0.5 1.0 1.5 Copenhagen University College of Engineering 2.0 Quotient 2.5 22 Two worlds of risk and uncertainty Uncertainty Imprecision Ignorance Lack of knowledge Statistical nature Randomness Variability Palisade 2012 Risk Conference World Representation and calculation Possibility Possibility distributions [a; …; b] Interval arithmetic Global optimisation Probability Probability distributions {µ; σ} Linear approximation Monte Carlo simulation Copenhagen University College of Engineering 23 Modelling by probability distributions The actual economic problem is modelled by a function Y of n independent and uncorrelated uncertain variables Y = Y(X1, X2,…, Xn). Linear approximation Y is approximated by means of a Taylor series Y Y(μ1,…, μn) + ∂Y/∂X1·(X1-μ1) + ∂Y/∂X2·(X2-μ2) + … + ∂Y/∂Xn·(Xn-μn), where ∂Y/∂Xi is the partial derivative of Y with respect to Xi calculated at (μ1,…, μn). The expected value is given by E(Y) = μ = Y(μ1,…, μn). The variance is approximated by VAR(Y) = σ2 (∂Y/∂X1)2·σ12 +…+ (∂Y/∂Xn)2·σn2. Palisade 2012 Risk Conference Copenhagen University College of Engineering Monte Carlo simulation 24 Modelling by possibility distributions i.e. intervals, fuzzy intervals, etc. The actual economic problem is modelled by a function Y of n uncertain variables Y = Y(X1, X2,…, Xn). NB: Function can be arranged in different ways. In case of intervals Y is calculated by means of interval arithmetic (only applicable in the simple case) or global optimisation (applicable in the general case). In case of triple estimates Extreme values of Y are calculated as above. In case of fuzzy intervals As above, for all α-cuts. Palisade 2012 Risk Conference Copenhagen University College of Engineering 25 Independent stochastic variables Intervals Triple estimates {μ; σ} = {μ1; σ1} # {μ2; σ2} [a; b] = [a1; b1] # [a2; b2] [a; c; b] = [a1; c1; b1] # [a2; c2; b2] μ = μ1 + μ2; a = a 1 + a 2; σ2 = σ12 + σ22 b = b1 + b2 μ = μ1 - μ2; a = a1 - b2; σ2 = σ12 + σ22 b = b1 - a 2 μ = μ1·μ2; a = min(a1a2, a1b2, b1a2, b1b2); σ2 σ12·μ22 + σ22·μ12 b = max(a1a2, a1b2, b1a2, b1b2) μ = μ1/μ2; a = min(a1/b2, a1/a2, b1/b2, b1/a2,); σ2 σ12/μ22 + σ22·μ12/μ24, b = max(a1/b2, a1/a2, b1/b2, b1/a2), Addition Subtraction Multiplication Division if μ2 ≠ 0 if 0 [a2; b2] a = a1 + a2; c = c1 + c2; b = b 1 + b2 a = a1 - b2; c = c1 - c2; b = b1 - a 2 a = min(a1a2, a1b2, b1a2, b1b2); c = c1c2; b = max(a1a2, a1b2, b1a2, b1b2) a = min(a1/b2, a1/a2, b1/b2, b1/a2,); c = c1/c2; b = max(a1/b2, a1/a2, b1/b2, b1/a2), if 0 [a2; b2] Table 1. Formulas for basic calculations with alternative representations of uncertain variables. Palisade 2012 Risk Conference Copenhagen University College of Engineering 26 Discounted cash flow case Net present value over n periods NPV = a0 + a1·(1+r1)-1 + a2·(1+r1)-1·(1+r2)-1 + …+ an·(1+r1)-1·(1+r2)-1·…·(1+rn)-1, ai = Xi1·Xi2 + Xi3 + Xi4 +…+ Xim, i = 0,…,n. ai: net cash flow in i’th period ri: rate of interest in i’th period Palisade 2012 Risk Conference Copenhagen University College of Engineering 27 Interval analysis ($1000) YEAR 0 YEAR 1 YEAR 2 YEAR 3 YEAR 4 [4.200; 5.200] [12.400; 14.100] [15.900; 18.100] [13.800; 15.600] Margin (%) [44,50; 45,50] % [45,00; 47,00] % [45,50; 48,50] % [44,00; 48,00] % Direct cost [-2.886; -2.289] [-7.755; -6.572] [-9.865; -8.188] [-8.736; -7.176] Margin [-1.869; 2.366] [5.580; 6.672] [7.234; 8.779] [6.072; 7.488] [-1.000; -800] [-975; -700] [-800; -600] [-800; -600] [-950; -700] [-1.375; -1.225] [-675; -525] [-675; -525] Turnover Marketing cost [-1.050; -950] Indirect production cost RD&E cost [-3.050; -2.950] [-1.700; -1.400] [-350; -250] [-150; -50] [-150; -50] Operating income [-4.100; -3.900] [-1.781; -534] [2.880; 4.452] [5.609; 7.604] [4.447; 6.313] Investment [-5.100; -4.900] [-2.200; -1.900] Net cash flow NCF [-9.200; -8.800] [-3.981; -2.434] [2.880; 4.452] [5.609; 7.604] [4.447; 7.013] [8,50; 9,50] % [9,00; 11,00] % [9,50; 12,50] % [10,50; 13,50] % [-3.669, -2.223] [2.369; 3.764] [4.102; 5.871] [2.865; 4.901] Rate of interest r (%) Discounted cash flow DCF [-9.200; -8.800] Net present value NPV [-3.532; 3.514] [0; 700] Table 2a. Discounted cash flow analysis by interval analysis (Interval Solver 2000, overall absolute and relative precision 10-6). Input variables in shaded cells. Palisade 2012 Risk Conference Copenhagen University College of Engineering 28 Uniform probability distributions ($1000) YEAR 0 YEAR 1 YEAR 2 YEAR 3 YEAR 4 {4.700; 289} {13.250; 491} {17.000; 635} {14.700; 520} Margin (%) {45,00; 0,29} % {46,00; 0,58} % {47,00; 0,87} % {46,00; 1,15} % Direct cost {-2.585; 160} {-7.155; 276} {-9.010; 368} {-7.938; 328} Margin {2.115; 131} {6.095; 239} {7.990; 333} {6.762; 293} {-900; 58} {-838; 79} {-700; 58} {-700; 58} {-825; 72} {-1.300; 43} {-600; 43} {-600; 43} Turnover Marketing cost {-1.000; 29} Indirect production cost RD&E cost {-3.000; 29} {-1.550; 87} {-300; 29} {-100; 29} {-100; 29} Operating income {-4.000; 41} {-1.160; 182} {3.657; 257} {6.590; 342} {5.362; 303} Investment {-5.000; 58} {-2.050; 87} Net cash flow NCF {-9.000; 71} {-3.210; 202} {3.657; 257} {6.590; 342} {5.712; 364} {9,00; 0,29} % {10,00; 0,58} % {11,00; 0,87} % {12,00; 0,87} % {-2.945; 184} {3.050; 215} {4.952; 262} {3.832; 251} Rate of interest r (%) Discounted cash flow DCF {-9.000; 71} Net present value NPV {-111; 470} {350; 202} Table 2b. Discounted cash flow analysis by stochastic variables, formulas (11) - (19). Input variables (shaded cells) are derived from uniform probability distributions corresponding to the interval input variables in Table 2a, however converted to the form {μ; σ}. Palisade 2012 Risk Conference Copenhagen University College of Engineering 29 Net present values Input Interval/Fuzzy (double and triple estimates) Monte Carlo: {-111; 468} Stochastic Uniform [-3.532; 3.514] {-111; 470} Triangular [-3.532; 1.317; 3.514] {313; 355} Monte Carlo: {314; 354} Palisade 2012 Risk Conference Copenhagen University College of Engineering 30 Comparisons Relative frequency 0,0010 0,0008 0,0006 Input: Triangular NPV: Normal µ = 313, σ = 355 (Table 3b) 1,2 1,0 0,8 Input: Uniform NPV: Normal µ = -111, σ = 470 (Table 2b) 0,6 0,0004 0,4 0,0002 Worst case (Table 2a & 3a) 0,2 Best case (Table 2a & 3a) 0,0000 0,0 -3000 -2000 -1000 0 1000 2000 3000 Net Present Value ($1000) Palisade 2012 Risk Conference Copenhagen University College of Engineering 31 Membership function Most possible case (Table 3a) 0,0012 Conclusions • Probability and possibility representations are different! • If no knowledge of probability distribution is available, use possibility representation • If additional statistical knowledge is available use probability distribution… • … but be aware of under estimation of risk! Palisade 2012 Risk Conference Copenhagen University College of Engineering 32 Thank You!