stochastic optimization with risk measures · stochastic optimization with risk measures ima new...

26
Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial and Systems Engineering University of Wisconsin-Madison August 8, 2016 Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 1 / 25

Upload: others

Post on 11-Jun-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Stochastic Optimization with Risk Measures

IMA New Directions Short Course on Mathematical OptimizationJim Luedtke

Department of Industrial and Systems EngineeringUniversity of Wisconsin-Madison

August 8, 2016

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 1 / 25

Page 2: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Motivating Application: Portfolio Optimization

Let’s make some money!

Given a set of assets I, what fraction of my portfolio should I investeach asset to maximize return?

Return on asset i ∈ I is random: Ri ≥ 0

Let xi = fraction of portfolio invested in asset i

Constraints: Invest all, and no short-selling∑i∈I xi = 1, x ≥ 0

What is optimal solution if we want to maximize expected value?

max E[∑

i∈I Rixi

]=∑

i∈I E[Ri]xi

Invest everything in one asset with highest expected return!

Anybody see a problem with that solution?

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 2 / 25

Page 3: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Mean-Variance Model

Markowitz key observation:

Variance of return is also important

For given expected return, want portfolio with smallest variance (orvice versa)

How to deal with two objectives?

Constrain expected return above L, minimize variance (vary L)

Constraint variance below L, maximize expected return (vary L)

X Maximize expected return less λ*variance (vary λ)

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 3 / 25

Page 4: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Mean-Variance Model

Random returns Ri, i ∈ I:

Expected return ri, i ∈ IVariance Σii, Covariance Σij

What is variance of return if xi is invested in asset i?

Recall V(Z) = E[(Z − µZ)2

], Cov(Z, Y ) = E[(Z − µZ)(Y − µY )]

E[(∑

i

Rixi −∑i

rixi)2]

=E[∑

i

∑j

(Ri − ri)(Rj − rj)xixjRiRj]

=∑i,j

Cov(Ri, Rj)xixj = x>Σx

where Σ is the variance-covariance matrix

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 4 / 25

Page 5: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Mean-Variance Model

Mean-variance model:

maxx>r︸︷︷︸

E[f(x,ξ)]

−λ x>Σx︸ ︷︷ ︸V[f(x,ξ)]

: e>x = 1, x ≥ 0

Σ 0⇒ Convex optimization model

This is one example of a mean-risk optimization model

Possible drawback: Variance is symmetric ⇒ Penalizes high returns aswell as low returns

If returns are joint normally distributed, portfolio return is normal ⇒FineBut is normal a reasonable distribution of return?

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 5 / 25

Page 6: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Value at Risk

Another popular quantification of risk used in finance

α Value at Risk (VaR):

VaRα(Z) = inft : P[Z ≤ t] ≥ α= inft : FZ(t) ≥ α

Constraint on Value at Risk is equivalent to a chance constraint:

VaRα(f(x, ξ)) ≤ 0⇔ P[f(x, ξ) ≤ 0] ≥ α.

NB: In portfolio problem, evaluate VaR of loss: f(x, ξ) = −R>x

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 6 / 25

Page 7: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Risk Measures More Generally

Classic stochastic programming model:

minE[f(x, ξ)] : x ∈ X

Replace expected value with a risk measure:

minρ[f(x, ξ)] : x ∈ X

where ρ is a risk measure:

Maps a random variable to a number

E.g., ρ(Z) = E[Z], ρ(Z) = E[Z] + λV(Z), ρ(Z) = E[Z] + λVaR(Z),ρ(Z) = E[u(Z)], where u is a “disutility function”

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 7 / 25

Page 8: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Average Value at Risk

Another popular risk measure, often called “Conditional Value at Risk”

Minimizing, so wish to measure/limit risk of large values

Definition

α Average value at risk (AVaR)

AVaRα(Z)def= min

y∈R

y +

1

1− αE[Z − y]+

Whaa?!?Theorem (Intuitive Definition)

If, P(Z = VaRα(Z)) = 0, then

AVaRα(Z) = E[Z | Z ≥ VaRα(Z)]

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 8 / 25

Page 9: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Proving Equivalent Definition

Lemma

Assume P(Z = VaRα(Z)) = 0. Then

E[Z | Z ≥ VaRα(Z)] = VaRα(Z) +1

1− αE[Z − VaRα(Z)]+.

AVaRα(Z)def= min

y∈R

y +

1

1− αE[Z − y]+

(*)

Using Lemma,AVaRα(Z) = E[Z | Z ≥ VaRα(Z)]

follows if we can show v := VaRα(Z) is a minimizer of (*).

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 9 / 25

Page 10: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Optimizing Average Value at Risk

Assume finite support distribution: P(ξ = ξs) = ps, s = 1, . . . , S

minxAVaRα(f(x, ξ)) : x ∈ X

= minx,y

y +

1

1− αE[f(x, ξ)− y]+ : x ∈ X, y ∈ R

= min

x,yy +

1

1− α

S∑s=1

psws

s.t. ws ≥ f(x, ξs)− y, s = 1, . . . , S

w ≥ 0, y ∈ R, x ∈ X

Convex program when X is convex and f(x, ξs) is convex for each s

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 10 / 25

Page 11: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Properties of Average Value at Risk

Assume (only for simplicity) probability space is finite: |Ω| = S.

Rndom variable Z ⇔ vector z ∈ RS .

AVaRα: Function from RS → R.

Theorem

ρ(·) = AVaRα(·) satisfies the following properties:

(A1) Convexity. β ∈ (0, 1), Z1 and Z2 rvs ⇒

ρ(βZ1 + (1− β)Z2) ≤ βρ(Z1) + (1− β)ρ(Z2)

(A2) Monotonicity. Z1 ≥ Z2 a.s. ⇒ ρ(Z1) ≥ ρ(Z2)

(A3) Translation invariance. If Z is a r.v. and a ∈ R, thenρ(Z + a) = ρ(Z) + a.

(A4) Positive homogeneity. ρ(tZ) = tρ(Z) ∀t > 0.

Reference: Rockafellar and Uryasev [2000]

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 11 / 25

Page 12: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Coherent Risk Measures

AVaRα is an example of a coherent risk measure

Definition

A risk measure ρ(·) is called coherent if it satisfies the following proper-ties:

(A1) Convexity. β ∈ (0, 1), Z1 and Z2 rvs ⇒

ρ(βZ1 + (1− β)Z2) ≤ βρ(Z1) + (1− β)ρ(Z2)

(A2) Monotonicity. Z1 ≥ Z2 a.s. ⇒ ρ(Z1) ≥ ρ(Z2)

(A3) Translation invariance. If Z is a r.v. and a ∈ R, thenρ(Z + a) = ρ(Z) + a.

(A4) Positive homogeneity. ρ(tZ) = tρ(Z) ∀t > 0.

References: Artzner et al. [1999], Ruszczynski and Shapiro [2006]

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 12 / 25

Page 13: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Distributional Robustness

Recall assumption: |Ω| = S

Distribution P⇔ vector µ in RS+ with∑

s µs = 1

Random variable Z ⇔ vector z ∈ RS

In practice, distribution µ of random variable is not known with certainty

Assume instead µ ∈ A, where A is a collection of distributions

Find minimum cost over worst case distribution

Distributionally Robust Stochastic Optimization

minx∈X

maxµ∈A

Eµ[f(x, ξ)]

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 13 / 25

Page 14: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Distributional Robustness ⇒ Coherent Risk Measures

Suppose we are given a family of distributions ALet ρA(Z) := maxµ∈A Eµ[Z] (= maxµ∈A

∑s µszs)

Then DRO problem is equivalent to:

minx∈X

ρA[f(x, ξ)]

Distributional robustness provides a means to define a risk measure

NB: Evaluating ρA(Z) is tractable ⇔ A is a tractable convex set

Theorem

Suppose e>µ = 1 and µ ≥ 0 for all µ ∈ A. Then ρA is a coherent riskmeasure.

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 14 / 25

Page 15: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Coherent Risk Measures ⇒ Distributional Robustness

Theorem

Suppose ρ is a coherent risk measure. Then there exists a convex set Asuch that for any random variable Z

ρ[Z] = maxµ∈A

Eµ[Z]

where for each µ ∈ A,∑

s µs = 1 and µ ≥ 0.

Example: Consider random variable with P(Z = zs) = ps, where p ∈ RS+and

∑s ps = 1

AVaRα(Z) = maxµ∈Aα

Eµ[Z]

where Aα = µ ∈ RS+ :∑

s µs = 1, µs ≤ ps/(1− α)

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 15 / 25

Page 16: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Proof Background

Given a function ρ : RS → R, its conjugate ρ∗ : RS → R is:

ρ∗(µ) = maxz∈RSµ>z − ρ(z)

Let A := µ ∈ RS : ρ∗(µ) < +∞ρ∗ is a convex function and A is a convex set

Theorem

If ρ is a convex function, then

ρ(z) = maxµ∈Aµ>z − ρ∗(µ).

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 16 / 25

Page 17: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Another Example: Mean α-Semi-deviation

Let α ≥ 1, and for λ ≥ 0 define:

ρλ[Z] = E[Z] + λE[(Z − E[Z])α+

]1/αFor discrete distribution with P(Z = zs) = ps, s = 1, . . . , S

ρλ[Z] = p>z + λ(∑

s

ps(zs −∑i

pizi)α+

)1/αTheorem

If λ ∈ [0, 1] then ρλ(·) is a coherent risk measure.

The dispersion part by itself is NOT a coherent risk measure (itviolates monotonicity).

Proof of monotonicity of ρλ(·) is nontrivial.

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 17 / 25

Page 18: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Properties of Other Risk Measures

Mean Variance: ρMV [Z] = E[Z] + λV[Z]

Convex? Yes

Monotone? No(!)

Translation invariant? Yes

Positively homogeneous? No

Mean Standard-deviation: ρMS [Z] = E[Z] + λV(Z)1/2

Convex? Yes

Monotone? No(!)

Translation invariant? Yes

Positively homogeneous? Yes

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 18 / 25

Page 19: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Nonmonotonicity Example

Suppose there are two outcomes: P(1) = 0.95, P(2) = 0.05

ω Z1 Z2

1 10 102 10 0

E 10 9.5V 0 4.75

V1/2 0 2.18

Z1 ≥ Z2 with probability 1, but mean variance and meanstandard-deviation measures would prefer Z1 for modest values of λ

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 19 / 25

Page 20: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Properties of Risk Measures

VaRα(Z):

Convex? No

Monotone? Yes

Translation invariant? Yes

Positively homogeneous? Yes, if P(Z ≥ 0) = 1

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 20 / 25

Page 21: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Nonconvexity of Value at Risk

Suppose there are three equally likely outcomes

ω Z1 Z2 12Z

1 + 12Z

2

1 300 0 1502 0 0 03 0 300 150

VaR0.6 0 0 150E 100 100 100

12VaR0.6(Z

1) + 12VaR0.6(Z

1) < VaR0.6(12Z

1 + 12Z

2)

Diversified “portfolio” looks worse, despite having same expectedvalue

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 21 / 25

Page 22: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Disutility Based Risk Measure

Let g : R→ R be a convex increasing function. Let

ρg[Z] = E[g(Z)]

Properties:

Convex? Yes

Montone? Yes

Translation invariant? No. (Unless g(z) ≡ z.)

Positively homogeneous? Yes, if g is positively homogeneous.

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 22 / 25

Page 23: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Modified Disutility Based Risk Measure

Let g : R→ R be a convex increasing function. Let

ρg[Z]def= inf

tE[Z + g(Z − t)]

Properties:

Convex? Yes

Monotone? Yes

Translation invariant? Yes!

Positively homgeneous? Yes, if g is positively homgeneous.

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 23 / 25

Page 24: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Optimizing Risk Measures

Now consider the problem:

minρ[f(x, ξ)] : x ∈ X

Theorem

Assume ρ is a convex and monotone risk function and f(x, ξs) is convexfor s = 1, . . . , S. Then φ(x) := ρ[f(x, ξ)] is a convex function of x.

Let h : Rn → RS , be defined by hs(z) = f(x, ξs).

h is a convex function. ρ is a convex and increasing function.

⇒ φ(x) = ρ(h(x)) is convex.

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 24 / 25

Page 25: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Optimizing Risk Measures

Now consider the problem:

minφ(x) = ρ[f(x, ξ)] : x ∈ X

Let ρ be a coherent risk measure with representation:

ρ(Z) = maxµ∈A

Eµ[Z]

Assume f(x, ξs) is convex for each s = 1, . . . , S.

Subgradients

Let x ∈ X and µ ∈ argmaxµ∈A∑

s µsf(x, ξs), and let ds be a sub-gradient of f(x, ξs) at x for s ∈ Ω. Then∑

s

µsds

is a subgradient of φ(x) at x.

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 25 / 25

Page 26: Stochastic Optimization with Risk Measures · Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial

Philippe Artzner, Freddy Delbaen, Jean-Marc Eber, and David Heath.Coherent measures of risk. Mathematical finance, 9(3):203–228, 1999.

R.T. Rockafellar and S. Uryasev. Optimization of conditional value-at-risk.Journal of Risk, 2:21–41, 2000.

A. Ruszczynski and A. Shapiro. Optimization of convex risk functions.Mathematics of operations research, 31(3):433–452, 2006.

Jim Luedtke (UW-Madison) Risk Measures Lecture Notes 25 / 25