【OR】Robust Optimization(2):Distributionally Robust Shorfall Risk学习笔记
Navigator
One stage stochastic programming
考虑如下one stage
随机规划(SR
)
min
x
E
P
[
f
(
x
,
ξ
)
]
s
.
t
.
x
∈
X
\min_x \mathbb{E}_\mathcal{P}[f(x, \xi)]\\ s.t. \quad x\in X
xminEP[f(x,ξ)]s.t.x∈X
其中
f
:
R
n
×
R
k
→
R
f: \mathbb{R}^n\times \mathbb{R}^k\to \mathbb{R}
f:Rn×Rk→R为一个连续函数
ξ
:
Ω
→
Ξ
\xi: \Omega\to \Xi
ξ:Ω→Ξ为定义在概率空间
(
Ω
,
F
,
P
)
(\Omega, \mathcal{F}, P)
(Ω,F,P)上的随机变量,且
Ξ
⊂
R
k
\Xi\subset\mathbb{R}^k
Ξ⊂Rk
E
p
[
⋅
]
\mathbb{E}_p[\cdot]
Ep[⋅]为期望算子
分布鲁棒的主要思想:由于根据样本推测出总体的分布是不确定的,设置分布不确定集合为
P
\mathcal{P}
P,且满足未知真实分布的某些性质.
将原模型转为DRO
问题
min
sup
P
∈
P
E
P
[
f
(
x
,
ξ
)
]
s
.
t
.
x
∈
X
\min\sup_{P\in\mathcal{P}}\mathbb{E}_P[f(x, \xi)]\\ s.t.\quad x\in X
minP∈PsupEP[f(x,ξ)]s.t.x∈X
where P \mathcal{P} P is a set distributions satisfying certain known properties of the unknown true distribution.
Construction of ambiguity sets
- Monment-type ambiguity set
P = { P : E P [ ξ ] = μ , E P [ ( ξ − μ ) ( ξ − μ ) T ] = Σ } \mathcal{P}=\{P:\mathbb{E}_P[\xi]=\mu, \mathbb{E}_P[(\xi-\mu)(\xi-\mu)^T]=\Sigma\} P={P:EP[ξ]=μ,EP[(ξ−μ)(ξ−μ)T]=Σ} - Distance-type ambiguity set
P = { P : D ( P , P 0 ) ≤ r } \mathcal{P}=\{P:\mathbb{D}(P, P_0)\leq r\} P={P:D(P,P0)≤r}
其中 D \mathbb{D} D为概率空间中的度量,可以为 ϕ \phi ϕ-divergence,Kantorovich metric
,Wasserstain metric
. - Mixture distribution
根据Kantorovich metric
可以定义出Hausdorff distance
Utility-based shortfall risk measure
基于效用的风险度量,对于
Z
∈
L
∞
Z\in L^\infty
Z∈L∞定义如下
(
S
R
)
S
R
l
,
λ
P
(
Z
)
:
=
inf
{
t
∈
R
:
E
P
[
l
(
−
Z
−
t
)
]
≤
λ
}
(SR)\quad SR_{l,\lambda}^P(Z):=\inf\{t\in\mathbb{R}:\mathbb{E}_P[l(-Z-t)]\leq \lambda\}
(SR)SRl,λP(Z):=inf{t∈R:EP[l(−Z−t)]≤λ}
l
:
R
→
R
l:\mathbb{R}\to\mathbb{R}
l:R→R is an increasing and non-constant convex loss function.
λ
\lambda
λ is a pre-specified level of loss.(预设值
)
Alternative characterization
S
R
l
,
λ
P
(
Z
)
=
inf
{
t
∈
R
:
Z
+
t
∈
A
P
}
SR_{l, \lambda}^P(Z)=\inf\{t\in\mathbb{R}: Z+t\in\mathcal{A}_P\}
SRl,λP(Z)=inf{t∈R:Z+t∈AP}
其中,
A
P
:
=
{
Z
∈
L
∞
:
E
P
[
l
(
−
Z
)
]
≤
λ
}
\mathcal{A}_P:=\{Z\in L^\infty: \mathbb{E}_P[l(-Z)]\leq \lambda\}
AP:={Z∈L∞:EP[l(−Z)]≤λ}称之为可接受集合(acceptance set
)
该模型表示对于资产集
Z
Z
Z,注入最少的资金
t
t
t,使得
Z
+
t
Z+t
Z+t在可接受集合中.
当概率分布未知时,可以定义出分布鲁棒形式(Distributionally robust SR
)
(
D
R
S
R
)
S
R
l
,
λ
P
(
Z
)
:
=
inf
{
t
∈
R
:
sup
P
∈
P
E
P
[
l
(
−
Z
−
t
)
]
≤
λ
}
(DRSR) \quad SR_{l, \lambda}^{\mathcal{P}}(Z):=\inf\{t\in\mathbb{R}:\sup_{P\in\mathcal{P}}\mathbb{E}_P[l(-Z-t)]\leq \lambda\}
(DRSR)SRl,λP(Z):=inf{t∈R:P∈PsupEP[l(−Z−t)]≤λ}
Reformulation of DRSRP
在设置假设条件下,可以将DRSRP
模型转为这样一个凸问题
Convergence of the optimal value and optimal solutions
Portfolio
测试样本外表现(
Out-of-sample performance
)
Reference
西南大学优化讲座 郭少艳副教授
Distributionally robust shortfall risk optimization model and its approximation