The Recent Financial Turmoil and Related Financial Engineering Research Problems1 Steven Kou Columbia University 1 Based on Two Papers: Heyde, Kou, and Peng (2007). What Is a Good External Risk Measure: Bridging the Gaps between Robustness, Subadditivity, Prospect Theory, and Insurance Risk Measures. Peng and Kou (2008). Default Clustering and Pricing of CDO’s Steven Kou Columbia University () 1 / 49 1 Overview of Financial Engineering 2 The Recent Financial Turmoil 3 A New Model for CDO’s What is a CDO? Current Portfolio Credit Risk Models The New Conditional Survival (CS) Model CDO Pricing under the CS Model 4 What are Good External Risk Measures? Review of Risk Measures Motivation and Examples Reasons to Relax Subadditivity New Axioms and Characterization of Natural Risk Statistic 5 Summary Steven Kou Columbia University () 2 / 49 Outline 1 Overview of Financial Engineering 2 The Recent Financial Turmoil 3 A New Model for CDO’s What is a CDO? Current Portfolio Credit Risk Models The New Conditional Survival (CS) Model CDO Pricing under the CS Model 4 What are Good External Risk Measures? Review of Risk Measures Motivation and Examples Reasons to Relax Subadditivity New Axioms and Characterization of Natural Risk Statistic 5 Summary Steven Kou Columbia University () 2 / 49 Financial Engineering Apply engineering modeling and tools to build realistic (i.e. sophisticated) …nancial models, and to get simple (analytical or approximate) solutions. Steven Kou Columbia University () 3 / 49 Financial Engineering Apply engineering modeling and tools to build realistic (i.e. sophisticated) …nancial models, and to get simple (analytical or approximate) solutions. Apply …nancial methods to solve engineering related problems. Electricity Options Revenue Management (arbitrage and airline revenue management) Paper in Management Science 2008. Steven Kou Columbia University () 3 / 49 Outline 1 Overview of Financial Engineering 2 The Recent Financial Turmoil 3 A New Model for CDO’s What is a CDO? Current Portfolio Credit Risk Models The New Conditional Survival (CS) Model CDO Pricing under the CS Model 4 What are Good External Risk Measures? Review of Risk Measures Motivation and Examples Reasons to Relax Subadditivity New Axioms and Characterization of Natural Risk Statistic 5 Summary Steven Kou Columbia University () 4 / 49 The Recent Financial Turmoil The crisis in subprime credit markets Clustering defaults across time (time series correlation, Das et al. 2007) and cross-sectionally (contagion e¤ect, Longsta¤ and Rajan 2007) Di¢ culties in modeling collateralized debt obligations (CDOs) Risk Measures and BASEL Accord regulation Steven Kou Columbia University () 5 / 49 Outline 1 Overview of Financial Engineering 2 The Recent Financial Turmoil 3 A New Model for CDO’s What is a CDO? Current Portfolio Credit Risk Models The New Conditional Survival (CS) Model CDO Pricing under the CS Model 4 What are Good External Risk Measures? Review of Risk Measures Motivation and Examples Reasons to Relax Subadditivity New Axioms and Characterization of Natural Risk Statistic 5 Summary Steven Kou Columbia University () 6 / 49 What is a CDO What is a CDO (Collateralized Debt Obligation)? CDO is a security constructed from a portfolio of …xed-income securities or credit derivatives. CDO provides a way to create high quality debt out of a portfolio of low quality debt. Steven Kou Columbia University () 7 / 49 Steven Kou Columbia University () 8 / 49 CDO and Current Subprime Mortgage Crisis The impact of credit crunch on iTraxx 5Y index tranche spreads Tranches(%) 09/20/07 03/14/08 09/12/08 Steven Kou Columbia University () 0-3 1812 5150 4011 3-6 84 649 475 6-9 37 401 277 9-12 23 255 152 12-22 15 143 68 22-100 7 70 35 8 / 49 CDO and Current Subprime Mortgage Crisis The impact of credit crunch on iTraxx 5Y index tranche spreads Tranches(%) 09/20/07 03/14/08 09/12/08 0-3 1812 5150 4011 3-6 84 649 475 6-9 37 401 277 9-12 23 255 152 12-22 15 143 68 22-100 7 70 35 CDO played an important role in the subprime mortgage crisis Rating Agencies Banks Steven Kou Columbia University () 8 / 49 CDO and Current Subprime Mortgage Crisis The impact of credit crunch on iTraxx 5Y index tranche spreads Tranches(%) 09/20/07 03/14/08 09/12/08 0-3 1812 5150 4011 3-6 84 649 475 6-9 37 401 277 9-12 23 255 152 12-22 15 143 68 22-100 7 70 35 CDO played an important role in the subprime mortgage crisis Rating Agencies Banks Mispricing and improper credit rating of CDO have been criticized. CDO itself should not be blamed for the …nancial crisis: it is a good instrument as long as one knows how to price it. Steven Kou Columbia University () 8 / 49 The Market Standard — Gaussian Copula Models Idea: using copula functions to model default time correlation Literature: Gaussian copula model (Li, 2000) What is wrong with Gaussian copula? Gaussian copula cannot generate tail dependence lim P (Y > VaRq (Y )jX > VaRq (X )) = 0 q !1 Gaussian copula does not work during crisis, when the default correlation is strong. Steven Kou Columbia University () 9 / 49 Implied Copula Correlation of iTraxx 5Y CDO on 03/14/08 implied copula correlation 1 0.8 0.6 0.4 0.2 0 0-3% 3-6% Steven Kou Columbia University () 6-9% 9-12% Tranche 12-22% 22-100% 10 / 49 Other Models Top-down approach builds models for the portfolio cumulative loss process directly Good …t for standard CDO portfolios, but with no connection to underlying single names Longsta¤ and Rajan, 2007, Giesecke, et al., 2007, Halperin, 2007, Cont and Minca, 2007 Bottom-up approach builds models for single name default times Consistent with single name CDS spreads, but is hard to price and calibrate CDO tranches Examples: Static bottom-up models, e.g. copula models, dynamic intensity models Steven Kou Columbia University () 11 / 49 Dynamic Bottom-up Models General idea: Instantaneous default intensity of a single name P (τ i t + ∆t jτ i > t ) = λi (t )∆t + o (∆t ) Building correlation among instantaneous intensities λ1 (t ), . . . , λn (t ) Steven Kou Columbia University () 12 / 49 Dynamic Bottom-up Models General idea: Instantaneous default intensity of a single name P (τ i t + ∆t jτ i > t ) = λi (t )∆t + o (∆t ) Building correlation among instantaneous intensities λ1 (t ), . . . , λn (t ) Similar to survivial analysis in biostatistics Steven Kou Columbia University () 12 / 49 Dynamic Bottom-up Models General idea: Instantaneous default intensity of a single name P (τ i t + ∆t jτ i > t ) = λi (t )∆t + o (∆t ) Building correlation among instantaneous intensities λ1 (t ), . . . , λn (t ) Similar to survivial analysis in biostatistics Factor instantaneous intensity models(Du¢ e and Gârleanu, 2001; Mortensen, 2006) λi (t ) = ai λM (t ) + λid i (t ) M id λ (t ) and λi (t ) are independent a¢ ne jump di¤usion processes Other papers: Papageorgiou and Sircar (2007), Joshi and Stacey (2006), Schönbucher (2007) Steven Kou Columbia University () 12 / 49 Motivation A good model should be able to produce clustering defaults, both crosstionally and across the time. Traditional intensity models τ i = inf ft Λi (t ) = Z t 0 0 : Λi (t ) λi (s )ds Ei g , Ei d exp(1), i.i.d. (Λi (t ) is continuous!) The continuity does not allow simultaneous defaults of several names. Steven Kou Columbia University () 13 / 49 The New Model: Conditional Survival (CS) Model Our new model: conditional survival (CS) model is based on cumulative intensity Steven Kou Columbia University () 14 / 49 The New Model: Conditional Survival (CS) Model Our new model: conditional survival (CS) model is based on cumulative intensity Steven Kou Columbia University () 14 / 49 The New Model: Conditional Survival (CS) Model Our new model: conditional survival (CS) model is based on cumulative intensity τ i = inf ft Λi (t ) = J 0 : Λi (t ) Ei g , Ei d exp(1), i.i.d. ∑ ai ,j Mj (t ) + Xiid (t ), i = 1, . . . , n. j =1 M (t ) = (M1 (t ), . . . , MJ (t )) 0 are market factor processes: increasing and right continuous (possibly having jumps). Xiid (t ) 0 is the idiosyncratic part of the cumulative default intensity of the i-th name. Steven Kou Columbia University () 14 / 49 The New Model: Conditional Survival (CS) Model The new model Can produce simultaneous defaults of several names which leads to stronger correlation. Can lead to clustering default across the time. No dynamics of idiosyncratic cumulative intensity is needed The model provides automatic and exact calibration to single name CDS data The sensitivities w.r.t. each of the n single CDS’s can be obtained concurrently with the CDO pricing Steven Kou Columbia University () 15 / 49 Conditional Survival Probability Conditional survival probability h qic (t ) , P ( τ i > t j M (t )) = E e X iid (t ) Survival probability h qi (t ) , P (τ i > t ) = E e Steven Kou Columbia University () X iid (t ) i i h E e e ∑Jj=1 a i ,j M j (t ) ∑Jj=1 a i ,j M j (t ) i 16 / 49 Conditional Survival Probability Conditional survival probability h qic (t ) , P ( τ i > t j M (t )) = E e X iid (t ) Survival probability h qi (t ) , P (τ i > t ) = E e X iid (t ) i i h E e e ∑Jj=1 a i ,j M j (t ) ∑Jj=1 a i ,j M j (t ) i Conditional survival probability is the building block qic (t ) = qi (t ) Steven Kou Columbia University () e h E e ∑Jj=1 a i ,j M j (t ) ∑Jj=1 a i ,j M j (t ) i 16 / 49 Conditional Survival Probability Conditional survival probability h qic (t ) , P ( τ i > t j M (t )) = E e X iid (t ) Survival probability h qi (t ) , P (τ i > t ) = E e X iid (t ) i i h E e e ∑Jj=1 a i ,j M j (t ) ∑Jj=1 a i ,j M j (t ) i Conditional survival probability is the building block qic (t ) = qi (t ) e h E e ∑Jj=1 a i ,j M j (t ) ∑Jj=1 a i ,j M j (t ) i No dynamics of idiosyncratic cumulative intensity Xiid (t ) Using survival probability qi (t ) as input, the model provides automatic and exact calibration to single name CDS data. Steven Kou Columbia University () 16 / 49 Specifying Dynamics of Market Factors A choice for M (t ) is Polya process A Polya process is a mixed Poisson process with the rate λ having Gamma (α, β) distribution Clustering jumps: A Polya process has positive correlation between increments, so the arrival of one event tends to trigger more events. Polya process with parameter (α, β) has a negative binomial distribution P (M (t ) = i ) = α+i i 1 1 1 + βt α βt 1 + βt i E [exp( aM (t ))] has closed form. Steven Kou Columbia University () 17 / 49 CDO Pricing by Exact Simulation Key fact: conditional on M (t ), default indicators Ii = 1fτi t g , i = 1, . . . , n are independent Bernoulli (1 Steven Kou Columbia University () qic (t )) r.v. 18 / 49 CDO Pricing by Exact Simulation Key fact: conditional on M (t ), default indicators Ii = 1fτi t g , i = 1, . . . , n are independent Bernoulli (1 qic (t )) r.v. Exact simulation of Lt at given time t: 1 Generate market factors M (t ), M (t ), . . . , M (t ). 1 2 J 2 Calculate the conditional survival probability analytically qic (t ) 3 4 e h = qi (t ) E e d Generate independent Ii s Bernoulli(1 Calculate Lt = ∑ni=1 (1 Ri )Ni Ii . Steven Kou Columbia University () ∑Jj=1 a i ,j M j (t ) ∑Jj=1 a i ,j M j (t ) i qic (t )), i = 1, . . . , n. 18 / 49 CDO Pricing by Exact Simulation Key fact: conditional on M (t ), default indicators Ii = 1fτi t g , i = 1, . . . , n are independent Bernoulli (1 qic (t )) r.v. Exact simulation of Lt at given time t: 1 Generate market factors M (t ), M (t ), . . . , M (t ). 1 2 J 2 Calculate the conditional survival probability analytically qic (t ) 3 4 e h = qi (t ) E e d Generate independent Ii s Bernoulli(1 Calculate Lt = ∑ni=1 (1 Ri )Ni Ii . ∑Jj=1 a i ,j M j (t ) ∑Jj=1 a i ,j M j (t ) i qic (t )), i = 1, . . . , n. It takes 20 seconds to carry out 100,000 replication, using C++ program and a desktop with an Intel 2.33G cpu. Steven Kou Columbia University () 18 / 49 CDO Pricing by Exact Simulation Key fact: conditional on M (t ), default indicators Ii = 1fτi t g , i = 1, . . . , n are independent Bernoulli (1 qic (t )) r.v. Exact simulation of Lt at given time t: 1 Generate market factors M (t ), M (t ), . . . , M (t ). 1 2 J 2 Calculate the conditional survival probability analytically qic (t ) 3 4 e h = qi (t ) E e d Generate independent Ii s Bernoulli(1 Calculate Lt = ∑ni=1 (1 Ri )Ni Ii . ∑Jj=1 a i ,j M j (t ) ∑Jj=1 a i ,j M j (t ) i qic (t )), i = 1, . . . , n. It takes 20 seconds to carry out 100,000 replication, using C++ program and a desktop with an Intel 2.33G cpu. Variance reduction by using control variant E [Lt ] = ∑ni=1 (1 Ri )Ni (1 qi (t )) Steven Kou Columbia University () 18 / 49 Sensitivity w.r.t. Single Name CDS The sensitivity w.r.t. to an individual CDS can be reduced to the sensitivity w.r.t. to single name survival probability In CS model, expected tranche loss is linear w.r.t. single name survival probability E [(Lt a)+ ] = E fA(t )qi (t ) + B (t )g , e A(t ) = h E e ∑Jj=1 a i ,j M j (t ) ∑Jj=1 a i ,j M j (t ) E [(Lt i i a)+ (Lt i + (1 ∂E [(Lt a)+ ] = E [A(t )] ∂qi (t ) Ri )Ni a)+ jM (t )] where Lt i is the portfolio loss excluding the i-th name. A(t ) can be easily calculated in each simulation trial for pricing. The sensitivities w.r.t. each of the n single name CDS are obtained concurrently with CDO tranche pricing. Steven Kou Columbia University () 19 / 49 iTraxx 5Y CDO Tranche Spread Calibration Results CDO and CDS data on March 14, 2008, right before the collapse of Bear Stern, whose stock dropped to $2.84 on March 17, 2008. Steven Kou Columbia University () 20 / 49 iTraxx 5Y CDO Tranche Spread Calibration Results CDO and CDS data on March 14, 2008, right before the collapse of Bear Stern, whose stock dropped to $2.84 on March 17, 2008. We use single name 5Y CDS spread to obtain marginal survival probability qi (t ), i = 1, . . . , 125 Steven Kou Columbia University () 20 / 49 iTraxx 5Y CDO Tranche Spread Calibration Results CDO and CDS data on March 14, 2008, right before the collapse of Bear Stern, whose stock dropped to $2.84 on March 17, 2008. We use single name 5Y CDS spread to obtain marginal survival probability qi (t ), i = 1, . . . , 125 Calibration results by using 1 Polya process and 2 analog to integral of CIR processes Tranches(%) Market Model B-A spread Steven Kou Columbia University () 0-3 5150 5230 158 3-6 649 665 24 6-9 401 374 25 9-12 255 250 20 12-22 143 165 12 22-100 70 67 3 20 / 49 iTraxx 5Y CDO Tranche Spread Calibration Results CDO and CDS data on March 14, 2008, right before the collapse of Bear Stern, whose stock dropped to $2.84 on March 17, 2008. We use single name 5Y CDS spread to obtain marginal survival probability qi (t ), i = 1, . . . , 125 Calibration results by using 1 Polya process and 2 analog to integral of CIR processes Tranches(%) Market Model B-A spread 0-3 5150 5230 158 3-6 649 665 24 6-9 401 374 25 9-12 255 250 20 12-22 143 165 12 22-100 70 67 3 Pricing error: Chi-square = 6.7(p-value = 0.24), RMSE = 1.01, v !2 u 6 o 2 o u1 6 s s ( sk s k ) k k CHISQ = ∑ , RMSE = t ∑ o,a o,b s 6 s s k k =1 k =1 k k Steven Kou Columbia University () 20 / 49 Model and Market Implied Correlation Implied Copula Correlation of iTraxx 5Y C DO Tranches Spread on 3/14/08 1 Market Model 0.9 implied copula correlation 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0-3% 3-6% 6-9% 9-12% 12-22% 22-100% tranche Steven Kou Columbia University () 21 / 49 Outline 1 Overview of Financial Engineering 2 The Recent Financial Turmoil 3 A New Model for CDO’s What is a CDO? Current Portfolio Credit Risk Models The New Conditional Survival (CS) Model CDO Pricing under the CS Model 4 What are Good External Risk Measures? Review of Risk Measures Motivation and Examples Reasons to Relax Subadditivity New Axioms and Characterization of Natural Risk Statistic 5 Summary Steven Kou Columbia University () 22 / 49 Measure of Risk Maps a R.V. to a Real Number Risk: future loss of a position–random variable X Measure of risk: mapping from a set of risks to the real line X ! ρ (X ) Examples: margin requirement for …nancial trading capital requirement against market risk insurance risk premium Steven Kou Columbia University () 23 / 49 Coherent Measure of Risk Coherent Measure of Risk (Artzner et al. 1999, Huber 1981) Translation invariance: ρ(X + a) = ρ(X ) + a Positive homogeneity: ρ(λX ) = λρ(X ), λ Monotonicity: ρ(X ) ρ(Y ), if X Subadditivity: ρ(X + Y ) Steven Kou Columbia University () 0 Y ρ (X ) + ρ (Y ) 24 / 49 Coherent Measure of Risk Coherent Measure of Risk (Artzner et al. 1999, Huber 1981) Translation invariance: ρ(X + a) = ρ(X ) + a Positive homogeneity: ρ(λX ) = λρ(X ), λ Monotonicity: ρ(X ) ρ(Y ), if X Subadditivity: ρ(X + Y ) 0 Y ρ (X ) + ρ (Y ) Characterization of coherent measure of risk Maximum expected loss ρ(X ) = sup E P [X ] P 2P Acceptance sets ρ(X ) = minfm : X Steven Kou Columbia University () m 2 acceptance setg 24 / 49 Insurance Measure of Risk X and Y are comonotonic if and only if 8 ω 1 , ω 2 2 Ω, (X (ω 1 ) Steven Kou Columbia University () X (ω 2 ))(Y (ω 1 ) Y (ω 2 )) 0 25 / 49 Insurance Measure of Risk X and Y are comonotonic if and only if 8 ω 1 , ω 2 2 Ω, (X (ω 1 ) X (ω 2 ))(Y (ω 1 ) Y (ω 2 )) 0 Insurance Measure of Risk (Wang et al. 1997) Law invariance: ρ(X ) = ρ(Y ), if X and Y have the same distribution Monotonicity: ρ(X ) ρ(Y ), if X Y Comonotonic additivity: ρ(X + Y ) = ρ(X ) + ρ(Y ), if X and Y are comonotonic Continuity: lim ρ(max(X d !0 d, 0)) = ρ(X + ), lim ρ(min(X , d )) = ρ(X ) lim ρ(max(X , d )) = ρ(X ) d !∞ d! ∞ Scale normalization: ρ(1) = 1 Steven Kou Columbia University () 25 / 49 Insurance Measure of Risk (Continued) It does not require subadditivity ρ( ) is an insurance measure of risk if and only if ρ( ) has a Choquet integral representation: ρ (X ) = = Z Xd (g Z 0 ∞ P) (g (P (X > t )) 1)dt + Z ∞ g (P (X > t ))dt 0 where g ( ) is increasing, g (0) = 0, g (1) = 1 It does not generate scenarios, unlike the coherent risk measure Steven Kou Columbia University () 26 / 49 What Is an Axiom De…nition: a statement or principle which is generally accepted to be true, but is not necessarily so. (Cambridge English Dictionary) Axioms are subject to debate and change. It is useful to provide alternative axioms. Steven Kou Columbia University () 27 / 49 What Is an Axiom De…nition: a statement or principle which is generally accepted to be true, but is not necessarily so. (Cambridge English Dictionary) Axioms are subject to debate and change. It is useful to provide alternative axioms. Whether a risk measure is good or not depends on the objective of proposing the risk measure A risk measure may be suitable for internal risk management, but not for external regulatory agencies, and vice versa. We focus on external risk measures from the viewpoint of governmental/regulatory agencies Steven Kou Columbia University () 27 / 49 Legal Realism and Legal Realism Legal Realism Legal Realism: The legal decision of a court is determined by the actual practices of judges, rather than the law set forth in statutes or precedents (Hart, 1994). A law is only a guideline for judges and enforcement o¢ cers The law should be robust. Coherent risk measure generally lack robustness Steven Kou Columbia University () 28 / 49 Legal Realism and Legal Realism Legal Realism Legal Realism: The legal decision of a court is determined by the actual practices of judges, rather than the law set forth in statutes or precedents (Hart, 1994). A law is only a guideline for judges and enforcement o¢ cers The law should be robust. Coherent risk measure generally lack robustness Legal Positivism Legal Positivism: the existence and content of law depend on social norms and not on their merits (Hart, 1994). If a system of rules are to be imposed by force, there must be a su¢ cient number of people who accept it voluntarily. A law should re‡ect the social norms. In face of …nancial risk, people’s decision may violate the subadditivity Axiom. (Kahneman and Tversky, 1979, 1992) Steven Kou Columbia University () 28 / 49 An Example of Tra¢ c Speed Limit National Maximum Speed Law was repealed in 1995 because of notoriously low compliance. Not by studying geometrical or architectural features of road “85th percentile rule” The 85th percentile speed manifests: A law should be robust A law should be consistent to social behaviors Steven Kou Columbia University () 29 / 49 Example: Value-at-risk (VaR) Value-at-risk (VaR) at level α: VaRα (X ) = α-quantile of X , α 2 [95%, 99.9%] Value-at-risk satis…es axioms for the insurance risk measure Value-at-risk does not satisfy subadditivity VaRα (X + Y ) VaRα (X ) + VaRα (Y ), in general Inconsistency: 99% VaR is the primary measure of risk imposed by Basel Accord, but it is excluded by coherent measure of risk Steven Kou Columbia University () 30 / 49 Example: Value-at-risk (VaR) Value-at-risk (VaR) at level α: VaRα (X ) = α-quantile of X , α 2 [95%, 99.9%] Value-at-risk satis…es axioms for the insurance risk measure Value-at-risk does not satisfy subadditivity VaRα (X + Y ) VaRα (X ) + VaRα (Y ), in general Inconsistency: 99% VaR is the primary measure of risk imposed by Basel Accord, but it is excluded by coherent measure of risk Steven Kou Columbia University () 30 / 49 Example: Tail Conditional Expectation (TCE) Tail conditional expectation (TCE) at level α (Artzner et.al, 1999): TCEα (X ) = E [X jX VaRα (X )] Tail conditional expectation is also called conditional value-at-risk (Rockafellar & Uryasev, 2002) and expected shortfall (Acerbi, 2002) Tail conditional expectation is subadditive. Tail conditional expectation is not robust. Steven Kou Columbia University () 31 / 49 Tail conditional median (TCM) A more robust risk measure: Tail conditional median (TCM) at level α: TCMα (X ) = median [X jX Steven Kou Columbia University () VaRα (X )] 32 / 49 Tail conditional median (TCM) A more robust risk measure: Tail conditional median (TCM) at level α: TCMα (X ) = median [X jX VaRα (X )] = VaR 1+α (X ) 2 Tail conditional median does NOT satisfy subadditivity The criticism that VaR is not informative is inappropriate. TCM satis…es subadditivity in examples in which VaR violates subadditivity. Steven Kou Columbia University () 32 / 49 TCE TCM 9 9 Laplace T-3 T-5 T-12 8 8 7 α 1.44 6 TCM TCE α 7 5 4 4 3 3 -6.5 -6 -5.5 log(1- α) Steven Kou Columbia University () -5 -4.5 0.75 6 5 -7 Laplace T-3 T-5 T-12 -7 -6.5 -6 -5.5 log(1- α) -5 -4.5 33 / 49 Regulator’s Viewpoint: Robustness Is Crucial When a regulator impose a risk measure: Each …rm calculates the level of risk and reports to the regulator Each …rm has the freedom to choose its own internal models The risk measure should be robust with respect to underlying models. Steven Kou Columbia University () 33 / 49 Regulator’s Viewpoint: Robustness Is Crucial When a regulator impose a risk measure: Each …rm calculates the level of risk and reports to the regulator Each …rm has the freedom to choose its own internal models The risk measure should be robust with respect to underlying models. Heaviness of tail distribution is hard to identify With 5,000 observations one can not distinguish between the Laplace distribution and the T-distributions. (Heyde & Kou, 2004) Steven Kou Columbia University () 33 / 49 Axioms for Natural Risk Statistic Data x̃ = (x1 , . . . , xn ) 2 Rn Risk statistic: a mapping from the data set to the real line: ρ̂ : Rn ! R x̃ 7! ρ̂(x̃ ) Axioms for natural risk statistic: Positive homogeneity and translation invariance: ρ̂(ax̃ + b ) = aρ̂(x̃ ) + b, 8x̃ 2 Rn , a 0, b 2 R Monotonicity: ρ̂(x̃ ) ρ̂(ỹ ), if x̃ ỹ , i.e., xi yi , i = 1, . . . , n. Comonotonic subadditivity: ρ̂(x̃ + ỹ ) ρ̂(x̃ ) + ρ̂(ỹ ), if x̃ and ỹ are comonotonic, i.e., (xi xj )(yi yj ) 0 for any i 6= j. Empirical law invariance: ρ̂((x1 , . . . , xn )) = ρ̂((xi1 , . . . , xin )), for any permutation (i1 , . . . , in ). Steven Kou Columbia University () 34 / 49 Representation Theorem for Natural Risk Statistic Theorem Let x(1 ) , ..., x(n ) be the order statistics of the data x̃ = (x1 , ..., xn ) with x(n ) being the largest. Then ρ̂ is a natural risk statistic if and only if there exists a set of weights W = fw̃ = (w1 , . . . , wn )g Rn with each w̃ 2 W satisfying ∑ni=1 wi = 1 and wi 0, 81 i n, such that n ρ̂(x̃ ) = sup f ∑ wi x(i ) g, 8x̃ 2 Rn . w̃ 2W i =1 Natural risk statistic includes VaR along with scenario analysis and TCM Steven Kou Columbia University () 35 / 49 Representation Theorem for Natural Risk Statistic Theorem Let x(1 ) , ..., x(n ) be the order statistics of the data x̃ = (x1 , ..., xn ) with x(n ) being the largest. Then ρ̂ is a natural risk statistic if and only if there exists a set of weights W = fw̃ = (w1 , . . . , wn )g Rn with each w̃ 2 W satisfying ∑ni=1 wi = 1 and wi 0, 81 i n, such that n ρ̂(x̃ ) = sup f ∑ wi x(i ) g, 8x̃ 2 Rn . w̃ 2W i =1 Natural risk statistic includes VaR along with scenario analysis and TCM The proof does not follow easily from Huber’s result in robust statistics. Subadditivity only holds for comonotonic sets, which are not open sets. Boundary points do not have as nice properties as interior points. Steven Kou Columbia University () 35 / 49 Outline 1 Overview of Financial Engineering 2 The Recent Financial Turmoil 3 A New Model for CDO’s What is a CDO? Current Portfolio Credit Risk Models The New Conditional Survival (CS) Model CDO Pricing under the CS Model 4 What are Good External Risk Measures? Review of Risk Measures Motivation and Examples Reasons to Relax Subadditivity New Axioms and Characterization of Natural Risk Statistic 5 Summary Steven Kou Columbia University () 36 / 49 Summary Steven Kou Columbia University () 37 / 49 Summary Steven Kou Columbia University () 37 / 49 Summary Steven Kou Columbia University () 37 / 49 The Future of Financial Engineering after the Recent Financial Turmoil Steven Kou Columbia University () 38 / 49 The Future of Financial Engineering after the Recent Financial Turmoil Olympic Slogan: “Citius, Altius, Fortius”– “Faster, Higher, Stronger” Steven Kou Columbia University () 38 / 49 The Future of Financial Engineering after the Recent Financial Turmoil Olympic Slogan: “Citius, Altius, Fortius”– “Faster, Higher, Stronger” Faster: Analytical Solutions and Approximations Steven Kou Columbia University () 38 / 49 The Future of Financial Engineering after the Recent Financial Turmoil Olympic Slogan: “Citius, Altius, Fortius”– “Faster, Higher, Stronger” Faster: Analytical Solutions and Approximations Higher: High dimension modeling in option pricing, credit risks, VaR, MBS Steven Kou Columbia University () 38 / 49 The Future of Financial Engineering after the Recent Financial Turmoil Olympic Slogan: “Citius, Altius, Fortius”– “Faster, Higher, Stronger” Faster: Analytical Solutions and Approximations Higher: High dimension modeling in option pricing, credit risks, VaR, MBS Stronger: better and more ‡exible models for stocks, interest rates, energy prices. Steven Kou Columbia University () 38 / 49 The Future of Financial Engineering after the Recent Financial Turmoil Olympic Slogan: “Citius, Altius, Fortius”– “Faster, Higher, Stronger” Faster: Analytical Solutions and Approximations Higher: High dimension modeling in option pricing, credit risks, VaR, MBS Stronger: better and more ‡exible models for stocks, interest rates, energy prices. Going from the Wall Street to the Main Street. Steven Kou Columbia University () 38 / 49 Thank you! Steven Kou Columbia University () 39 / 49 Implicit Constraints on Model Parameters The idiosyncratic cumulative intensities Xiid (t ) h i h i h id id E e X i (T m ) E e X i (T m 1 ) E e Recall that h E e X iid (t ) i h = E e 0 and increasing i X iid (T 1 ) 1, 81 i qi (t ) ∑Jj=1 a i ,j M j (t ) This imposes parameter constraints: h E e qi (Tm ) ∑Jj=1 a i ,j M j (T m ) Steven Kou Columbia University () i h E e qi (T1 ) ∑Jj=1 a i ,j M j (T 1 ) i i 1, 81 i n 40 / 49 n Specifying Dynamics of Market Factors (Continued) Choice B: model M (t ) by an anology to M (t ) = Rt 0 V (s )ds, V (t ) is a CIR process: K 1 1 1 V (t0 )h + ∑ V (ti )h + V (tK )h, 0 = t0 < t1 < 2 2 i =1 < tK = t h is the discretization step size, e.g. h = 1/32. Rt The exact simulation of 0 V (s )ds is time-consuming. The exact simulation of M (t ) is reasonably fast: only involves simulation of CIR process. E [exp( aM (t ))] has closed form. Steven Kou Columbia University () 41 / 49 Calibration Algorithm 1 2 Initialization: set market factor parameter Θ0 , and set s = 0. Iteration: s ! s + 1 For given Θs , determine loading coe¢ cients ai by solving a constrained optimization problem: i h J min E e ∑j =1 ai ,j N j (T m ) /qi (Tm ) 1 s.t. E e 0 q i (T m ) ∑Jj=1 a i ,j N j (T m ) E e q i (T 1 ) ∑Jj=1 a i ,j N j (T 1 ) 1 ai ,j These problems can be solved quickly by e.g., sequential quadratic programming(we used the package CFSQP developed by Lawence et.al). Calculate tranche spreads using the market parameter Θs and loading coe¢ cients ai obtained above. Update the market factor parameter Θs ! Θs +1 by using some optimization routine, e.g. Powell’s direction-set algorithm(Powell, 1964) Repeat Steven Kou Columbia University () 42 / 49 iTraxx 5Y CDO Tranche Spread Calibration Results Fitted model parameters: 1 2 3 Polya process: α = 2.64516, β = 0.00583919 1st CIR process: θ 1 = 0.1, κ 1 = 0.12792, σ1 = 1.34701, V1 (0) = 1.1411; h = 1/32 2nd CIR process: θ 2 = 0.1, κ 2 = 0.00100369, σ2 = 7.40483, V2 (0) = 0.106509; h = 1/32 Steven Kou Columbia University () 43 / 49 A merger may create extra risk X and Y : the net payo¤ of two …rms before a merger. Because of bankruptcy protection, the actual net payo¤ of the two …rms would be X + = max(X , 0) and Y + = max(Y , 0) After the merger, the net payo¤ of the joint …rm would be X + Y , and the actual net payo¤ would be (X + Y )+ , due to bankruptcy protection. (X + Y ) + Steven Kou Columbia University () X + + Y +! 44 / 49 Comonotonic Additivity Does Not Hold in Two Scenarios Probability space: (Ω = fω 1 , ω 2 , ω 3 g, F , P ), P (ω i ) = 1 3 Two comonotonic random variables (Z (ω 1 ), Z (ω 2 ), Z (ω 3 )) = (3, 2, 4), Y = Z 2 Two distortion function 1 2 g1 ( ) = 0.5, g1 ( ) = 1.0; 3 3 1 2 g2 ( ) = 0.72, g2 ( ) = 0.8 3 3 Risk measureRincorporating two R scenarios: ρ(X ) = max X d (g1 P ), X d (g2 P ) Only strict comonotonic subadditivity holds! ρ(Z + Y ) = 9.28 < ρ(Z ) + ρ(Y ) = 2.5 + 6.8 = 9.3 Steven Kou Columbia University () 45 / 49 1. Diversi…cation and Tail subadditivity of VaR For risks with …nite second moment, diversi…cation is preferable because SD (X + Y ) SD (X ) + SD (Y ) Diversi…ed portfolio has larger expected utility than a single risk. (Samuelson, 1967) Diversi…cation is not preferable if the risks have extremely heavy tails. (Ibragimov and Walden, 2004) VaR has tail subadditivity(Daníelsson et al. 2005): If the joint distribution of X and Y does not have extremely heavy tail, then there exists α0 2 (0, 1), such that VaRα (X + Y ) Diversi…cation reduces risk? VaR satis…es subadditivity? Steven Kou Columbia University () VaRα (X ) + VaRα (Y ), 8α 2 (α0 , 1) Not very heavy tails Yes Yes Heavy tails No No 46 / 49 Other Reasons to Relax Subadditivity Merger of two …rms removes the bankruptcy protection between them and hence increase the risk of complete breakdown of the joint …rm. (Dhaene et al. 2005) Prospect theory (Kahneman & Tversky, 1979, 1992, 1993) is able to explain the major violations of expected utility theory Non-comonotonic random loss can violate subadditivity Steven Kou Columbia University () 47 / 49 An Example Violating Subadditivity A fair coin is ‡ipped Three positions: H: losing $10,000 if it comes up “head" T: losing $10,000 if it comes up “tail" S: losing $5,000 for sure ρ(S ) = 5, 000, ρ(H ) = ρ(T ) ρ(H ) < ρ(S ) = 5, 000 because people are risk seeking in face of loss ρ(H + T ) = ρ(sure loss of 10,000) = 10, 000 Subadditivity does not hold: ρ(H ) + ρ(T ) < 10, 000 = ρ(H + T )! Steven Kou Columbia University () 48 / 49 Representation Theorem for Law Invariant Coherent Risk Statistic Theorem ρ̂ is a law invariant coherent risk statistic if and only if there exists a set of weights W = fw̃ = (w1 , . . . , wn )g Rn with each w̃ 2 W satisfying 0, 81 i n and w1 w2 . . . wn , such that ∑ni=1 wi = 1 and wi n ρ̂(x̃ ) = sup f ∑ wi x(i ) g, 8x̃ 2 Rn . w̃ 2W i =1 Steven Kou Columbia University () 49 / 49 Representation Theorem for Law Invariant Coherent Risk Statistic Theorem ρ̂ is a law invariant coherent risk statistic if and only if there exists a set of weights W = fw̃ = (w1 , . . . , wn )g Rn with each w̃ 2 W satisfying 0, 81 i n and w1 w2 . . . wn , such that ∑ni=1 wi = 1 and wi n ρ̂(x̃ ) = sup f ∑ wi x(i ) g, 8x̃ 2 Rn . w̃ 2W i =1 Assigning larger weights to larger observations leads to less robust risk statistics Coherent risk statistic is generally not robust! Steven Kou Columbia University () 49 / 49