CQF笔记M2L2优化理论及其在资产组合选择中的应用
Module 2 Quantitative Risk & Return
Lecture 2 Fundamentals of Optimization and Application to Portfolio Selectio
2.2.1 优化理论
优化问题:寻找函数 f f f满足一系列约束条件时的"最优可能值"
优化技术在金融领域有广泛应用:
- 计算债券收益的价值
- 求解资产组合选择问题(离散时间和连续时间)
- 衍生品估值(美式期权和Passport期权)
优化问题的形式多样,有的非常简单,有的极其复杂。
优化问题的一般形式:
min
x
1
,
⋯
,
x
n
f
(
x
1
,
⋯
,
x
n
)
\begin{aligned} \min_{x_1, \cdots, x_n} f(x_1,\cdots, x_n) \end{aligned}
x1,⋯,xnminf(x1,⋯,xn)
subject to :
g
1
(
x
1
,
⋯
,
x
n
)
≤
b
1
⋮
=
⋮
g
m
(
x
1
,
⋯
,
x
n
)
≥
b
m
\begin{aligned} & g_1(x_1,\cdots, x_n) & \le & \ b_1 \\ & \vdots &= & \ \vdots \\ & g_m(x_1,\cdots, x_n) & \ge & \ b_m \\ \end{aligned}
g1(x1,⋯,xn)⋮gm(x1,⋯,xn)≤=≥ b1 ⋮ bm
- f f f 目标函数
- x 1 , ⋯ , x n x_1,\cdots, x_n x1,⋯,xn 决策变量(decision variables)
- g 1 , ⋯ , g m g_1, \cdots, g_m g1,⋯,gm 约束条件,约束条件可以是等式也可以是不等式
基本规则
-
转换最大值和最小值
m a x f ( x ) = − m i n ( − f ( x ) ) m i n f ( x ) = − m a x ( − f ( x ) ) \begin{aligned} maxf(x) = -min(-f(x)) \\ minf(x) = -max(-f(x)) \\ \end{aligned} maxf(x)=−min(−f(x))minf(x)=−max(−f(x)) -
仿射变换(Affine Transformation)
m a x ( a + b f ( x ) ) = a + b m a x ( f ( x ) ) , b > 0 \begin{aligned} max(a+bf(x)) = a + b \ max(f(x)), \ b > 0 \end{aligned} max(a+bf(x))=a+b max(f(x)), b>0
2.2.2 无约束条件的优化问题
无约束条件的优化问题退化为 min x 1 , ⋯ , x n f ( x 1 , ⋯ , x n ) \begin{aligned} \min_{x_1, \cdots, x_n} f(x_1,\cdots, x_n) \end{aligned} x1,⋯,xnminf(x1,⋯,xn)
这是个标准的多元微积分问题,目标是找到一个向量 x ⃗ ∗ = ( x 1 ∗ , ⋯ , x n ∗ ) \vec{x}^* = (x_1^*, \cdots, x_n^*) x∗=(x1∗,⋯,xn∗)使得 f f f为全局最小值
- 必要条件: f f f在 x ⃗ ∗ \vec{x}^* x∗处的梯度(微分向量),记做 ∇ f ( x ⃗ ∗ ) \nabla f(\vec{x}^*) ∇f(x∗),为0,表示在 x ⃗ ∗ \vec{x}^* x∗处为极值,但是可能是极小值也可能是极大值
- 充分条件: f f f在 x ⃗ ∗ \vec{x}^* x∗处的海森矩阵(Hessian, 二阶微分矩阵),记做 H f ( x ⃗ ∗ ) Hf(\vec{x}^*) Hf(x∗),为正(为负表示极大值)
应用1 均值-方差优化准则
符号
单位向量
1 ⃗ = ( 1 ⋮ 1 ⋮ 1 ) \vec{1} = \left( \begin{matrix} 1 \\ \vdots \\ 1 \\ \vdots \\ 1 \\ \end{matrix} \right) 1=⎝⎜⎜⎜⎜⎜⎜⎛1⋮1⋮1⎠⎟⎟⎟⎟⎟⎟⎞
收益的期望值(期望收益)
μ ⃗ = ( μ 1 ⋮ μ i ⋮ μ n ) \vec{\mu} = \left( \begin{matrix} \mu _1 \\ \vdots \\ \mu _i \\ \vdots \\ \mu _n \\ \end{matrix} \right) μ=⎝⎜⎜⎜⎜⎜⎜⎛μ1⋮μi⋮μn⎠⎟⎟⎟⎟⎟⎟⎞
期望的标准差(向量)
σ ⃗ = ( σ 1 ⋮ σ i ⋮ σ n ) \vec{\sigma} = \left( \begin{matrix} \sigma _1 \\ \vdots \\ \sigma _i \\ \vdots \\ \sigma _n \\ \end{matrix} \right) σ=⎝⎜⎜⎜⎜⎜⎜⎛σ1⋮σi⋮σn⎠⎟⎟⎟⎟⎟⎟⎞
期望的标准差(对角阵)
S ⃗ = D ( σ ⃗ ) = ( σ 1 0 ⋯ 0 0 σ 2 ⋯ 0 ⋮ ⋮ ⋱ ⋮ 0 0 ⋯ σ n ) \vec{S} = {D}(\vec{\sigma}) = \left( \begin{matrix} \sigma_1 & 0 & \cdots & 0 \\ 0 & \sigma_2 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & \sigma_n \\ \end{matrix} \right) S=D(σ)=⎝⎜⎜⎜⎛σ10⋮00σ2⋮0⋯⋯⋱⋯00⋮σn⎠⎟⎟⎟⎞
期望的相关性矩阵(对称阵)
R ⃗ = ( 1 ρ 12 ⋯ ρ 1 n ρ 21 1 ⋯ ρ 2 n ⋮ ⋮ ⋱ ⋮ ρ n 1 ρ n 2 ⋯ 1 ) \vec{R} = \left( \begin{matrix} 1 & \rho_{12} & \cdots & \rho_{1n} \\ \rho_{21} & 1 & \cdots & \rho_{2n} \\ \vdots & \vdots & \ddots & \vdots \\ \rho_{n1} & \rho_{n2} & \cdots & 1 \\ \end{matrix} \right) R=⎝⎜⎜⎜⎛1ρ21⋮ρn1ρ121⋮ρn2⋯⋯⋱⋯ρ1nρ2n⋮1⎠⎟⎟⎟⎞
期望的协方差矩阵(对称阵)
Σ ⃗ = S R S ⃗ = ( σ 1 2 ρ 12 σ 1 σ 2 ⋯ ρ 1 n σ 1 σ n ρ 21 σ 2 σ 1 σ 2 2 ⋯ ρ 2 n σ 2 σ n ⋮ ⋮ ⋱ ⋮ ρ n 1 σ 2 σ n ρ n 2 σ n σ 2 ⋯ σ n 2 ) \vec{\Sigma} = \vec{SRS} = \left( \begin{matrix} \sigma_1^2 & \rho_{12} \sigma_1 \sigma_2 & \cdots & \rho_{1n} \sigma_1 \sigma_n \\ \rho_{21} \sigma_2 \sigma_1 & \sigma_2^2 & \cdots & \rho_{2n} \sigma_2 \sigma_n \\ \vdots & \vdots & \ddots & \vdots \\ \rho_{n1} \sigma_2 \sigma_n & \rho_{n2} \sigma_n \sigma_2 & \cdots & \sigma_n^2 \\ \end{matrix} \right) Σ=SRS=⎝⎜⎜⎜⎛σ12ρ21σ2σ1⋮ρn1σ2σnρ12σ1σ2σ22⋮ρn2σnσ2⋯⋯⋱⋯ρ1nσ1σnρ2nσ2σn⋮σn2⎠⎟⎟⎟⎞
资产权重
w ⃗ = ( w 1 ⋮ w i ⋮ w n ) \vec{w} = \left( \begin{matrix} w _1 \\ \vdots \\ w _i \\ \vdots \\ w _n \\ \end{matrix} \right) w=⎝⎜⎜⎜⎜⎜⎜⎛w1⋮wi⋮wn⎠⎟⎟⎟⎟⎟⎟⎞
投资组合中有N个资产,资产收益的
- 期望值为 μ ⃗ \vec{\mu} μ
- 标准差为 σ ⃗ \vec{\sigma} σ
- 相关系数为 R ⃗ \vec{R} R
- 权重分别为 ω ⃗ \vec{\omega} ω
组合收益的期望值
μ
π
=
μ
⃗
T
w
⃗
\mu_\pi = \vec{\mu} ^ T \vec{w}
μπ=μTw
组合收益的方差
σ
π
2
=
w
⃗
T
Σ
⃗
w
⃗
\sigma_\pi ^2 = \vec{w}^T \vec{\Sigma} \vec{w}
σπ2=wTΣw
注:这里按一般习惯用上标T表示向量/矩阵的转置
分解协方差矩阵
由于协方差矩阵 S ⃗ \vec{S} S是对角阵, 有 S ⃗ T = S ⃗ \vec{S}^T = \vec{S} ST=S,相关性矩阵 R ⃗ \vec{R} R是对称阵,也有 R ⃗ T = R ⃗ \vec{R}^T = \vec{R} RT=R,所以 Σ ⃗ T = S ⃗ T R ⃗ T S ⃗ = S ⃗ R ⃗ S ⃗ T = Σ ⃗ \vec{\Sigma}^T = \vec{S}^T \vec{R}^T \vec{S} = \vec{S} \vec{R} \vec{S}^T = \vec{\Sigma} ΣT=STRTS=SRST=Σ
加入无风险资产
无风险资产,收益的期望、收益的标准差和权重分别为 r , 0 , w 0 r, 0, w_0 r,0,w0
维度为n的单位向量记做
1
⃗
\vec{1}
1
1
⃗
=
(
1
⋮
1
⋮
1
)
\vec{1} = \left( \begin{matrix} 1 \\ \vdots \\ 1 \\ \vdots \\ 1 \\ \end{matrix} \right)
1=⎝⎜⎜⎜⎜⎜⎜⎛1⋮1⋮1⎠⎟⎟⎟⎟⎟⎟⎞
无风险资产的权重 w 0 = 1 − w ⃗ T 1 ⃗ w_0 = 1 - \vec{w}^T \vec{1} w0=1−wT1
组合收益的期望值
μ
π
=
μ
⃗
T
w
⃗
+
r
(
1
−
w
⃗
T
1
⃗
)
=
r
+
w
⃗
T
(
μ
⃗
−
r
1
⃗
)
\mu_\pi = \vec{\mu} ^ T \vec{w} + \ r \ (1 - \vec{w}^T \vec{1}) = r + \vec{w}^T (\vec{\mu} - r \vec{1})
μπ=μTw+ r (1−wT1)=r+wT(μ−r1)
引入无风险资产后,组合收益分为无风险收益
r
r
r和风险溢价
(
μ
⃗
−
r
1
⃗
)
(\vec{\mu} - r \vec{1})
(μ−r1)
组合收益的方差 σ π 2 = w ⃗ T Σ ⃗ w ⃗ \sigma_\pi ^2 = \vec{w}^T \vec{\Sigma} \vec{w} σπ2=wTΣw
资产组合优化的目标是找到权重向量 w ⃗ \vec{w} w,使得
- 收益一定时,风险(方差)最小
- 风险(方差)一定时,收益最大
马科维茨定义的目标函数(将风险和收益统一为效用函数):
max
w
μ
p
−
λ
2
σ
p
2
\begin{aligned} \max_w \mu_p - \frac{\lambda}{2} \sigma_p ^2 \end{aligned}
wmaxμp−2λσp2
其中
λ
\lambda
λ表示风险厌恶程度
由于将剩余的财产投入到无风险资产,不需要约束项
w
T
1
⃗
=
1
w^T \vec{1} = 1
wT1=1
因此优化问题变为: max w V ( w ) = [ r + w ⃗ T ( μ − r 1 ⃗ ) ] − λ 2 w ⃗ T Σ ⃗ w ⃗ \begin{aligned} \max_w V(w) = [r + \vec{w}^T (\mu - r \vec{1})] - \frac{\lambda}{2} \vec{w}^T \vec{\Sigma} \vec{w} \end{aligned} wmaxV(w)=[r+wT(μ−r1)]−2λwTΣw
一阶条件(梯度)
∇
V
(
w
⃗
∗
)
=
∂
V
∂
w
⃗
(
w
⃗
∗
)
=
(
μ
⃗
−
r
1
⃗
)
−
λ
Σ
⃗
w
⃗
=
0
\begin{aligned} \nabla V(\vec{w}^*) = \frac{\partial V}{\partial \vec{w}}(\vec{w}^*) = (\vec{\mu} - r \vec{1}) - \lambda \vec{\Sigma} \vec{w} = 0 \end{aligned}
∇V(w∗)=∂w∂V(w∗)=(μ−r1)−λΣw=0
得到备选解
w
∗
=
1
λ
Σ
⃗
−
1
(
μ
⃗
−
r
1
⃗
)
\begin{aligned} w^* = \frac{1}{\lambda} \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1}) \end{aligned}
w∗=λ1Σ−1(μ−r1)
海森矩阵
H
V
(
w
∗
)
=
−
λ
Σ
⃗
<
0
\begin{aligned} H V(w^*) = - \lambda \vec{\Sigma} < 0 \end{aligned}
HV(w∗)=−λΣ<0
因此备选解对应的是最大值
应用2 最小二乘法线性回归
CAPM 模型 : E [ R A − r ] = β E [ R M − r ] \mathbb{E}[R^A - r] = \beta \mathbb{E}[R^M - r] E[RA−r]=βE[RM−r]
- R A R^A RA 资产A的收益
- r r r 无风险利率
- β \beta β 对系统风险的敞口
- R M R^M RM 整个金融市场的收益
变换形式: E [ R A ] = r + β E [ R M − r ] \mathbb{E}[R_A] = r + \beta \mathbb{E}[R_M - r] E[RA]=r+βE[RM−r]
用最小二乘法(Ordinary Least Square, OLS)拟合历史数据,得到斜率 β \beta β的估计值 β ^ \hat{\beta} β^
- 对于成熟的公司,用5年的月度数据(60个观察值)
- 对于快速变化的公司,用两年的周数据(104个观察值)
R i A − r = β ( R i M − r ) + ϵ i R_i^A - r = \beta (R_i^M - r) + \epsilon_i RiA−r=β(RiM−r)+ϵi
- 对所有的 i i i, r r r和 β \beta β都是相同的,
-
ϵ
i
\epsilon_i
ϵi是每个历史数据的残差项error term,具备以下性质(为了数学推导上的方便)
- 均值为0: E [ ϵ i ] = 0 \mathbb{E}[\epsilon_i] = 0 E[ϵi]=0
- 方差为有限值: E [ ϵ i 2 ] = s 2 \mathbb{E}[\epsilon_i^2] = s^2 E[ϵi2]=s2
- 协方差为0: E [ ϵ i ϵ j ] = 0 \mathbb{E}[\epsilon_i \epsilon_j] = 0 E[ϵiϵj]=0
这里没有要求残差项服从正态分布,但是一般来说假设残差项服从正态分布会比较方便。
因变量
Y
⃗
=
(
Y
1
⋮
Y
i
⋮
Y
n
)
=
(
R
1
A
−
r
⋮
R
i
A
−
r
⋮
R
n
A
−
r
)
\vec{Y} = \left( \begin{matrix} Y _1 \\ \vdots \\ Y _i \\ \vdots \\ Y _n \\ \end{matrix} \right) = \left( \begin{matrix} R_1^A - r \\ \vdots \\ R_i^A - r \\ \vdots \\ R_n^A - r \\ \end{matrix} \right)
Y=⎝⎜⎜⎜⎜⎜⎜⎛Y1⋮Yi⋮Yn⎠⎟⎟⎟⎟⎟⎟⎞=⎝⎜⎜⎜⎜⎜⎜⎛R1A−r⋮RiA−r⋮RnA−r⎠⎟⎟⎟⎟⎟⎟⎞
自变量
X
⃗
=
(
X
1
⋮
X
i
⋮
X
n
)
=
(
R
1
M
−
r
⋮
R
i
M
−
r
⋮
R
n
M
−
r
)
\vec{X} = \left( \begin{matrix} X _1 \\ \vdots \\ X _i \\ \vdots \\ X _n \\ \end{matrix} \right) = \left( \begin{matrix} R_1^M - r \\ \vdots \\ R_i^M - r \\ \vdots \\ R_n^M - r \\ \end{matrix} \right)
X=⎝⎜⎜⎜⎜⎜⎜⎛X1⋮Xi⋮Xn⎠⎟⎟⎟⎟⎟⎟⎞=⎝⎜⎜⎜⎜⎜⎜⎛R1M−r⋮RiM−r⋮RnM−r⎠⎟⎟⎟⎟⎟⎟⎞
模型变为 Y ⃗ = X ⃗ β + ϵ ⃗ \vec{Y} = \vec{X} \beta + \vec{\epsilon} Y=Xβ+ϵ
ϵ ⃗ = ( ϵ 1 ⋮ ϵ i ⋮ ϵ n ) \vec{\epsilon} = \left( \begin{matrix} \epsilon _1 \\ \vdots \\ \epsilon _i \\ \vdots \\ \epsilon _n \\ \end{matrix} \right) ϵ=⎝⎜⎜⎜⎜⎜⎜⎛ϵ1⋮ϵi⋮ϵn⎠⎟⎟⎟⎟⎟⎟⎞
回归线方程 Y ⃗ ^ = X ⃗ β ^ \hat{\vec{Y}} = \vec{X} \hat{\beta} Y^=Xβ^
- β ^ \hat{\beta} β^是斜率 β \beta β的估计值
- Y ⃗ ^ \hat{\vec{Y}} Y^是 Y ⃗ \vec{Y} Y的预测值
残差项
ϵ
⃗
=
Y
⃗
−
X
⃗
β
^
=
Y
⃗
−
Y
⃗
^
\vec{\epsilon} = \vec{Y} - \vec{X} \hat{\beta} = \vec{Y} - \hat{\vec{Y}}
ϵ=Y−Xβ^=Y−Y^
估计值 β ^ \hat{\beta} β^应使得残差项最小,使用残差项的平方和作为目标函数
- 残差项项正可负,求平方可以消除负值的影响
- 求平方可以放大大的残差
- 求平方和可以累积所有观察值的影响
min β F ( β ) = ∑ i = 1 n ϵ i 2 = ∑ i = 1 n ( Y i − Y i ^ ) 2 = ∑ i = 1 n ( Y i − X i β ) 2 \min_{\beta} F(\beta) = \sum_{i=1}^{n} \epsilon_i^2 =\sum_{i=1}^{n} (Y_i - \hat{Y_i})^2 = \sum_{i=1}^{n} (Y_i - X_i \beta)^2 βminF(β)=i=1∑nϵi2=i=1∑n(Yi−Yi^)2=i=1∑n(Yi−Xiβ)2
矩阵表达形式
min
β
F
(
β
)
=
ϵ
⃗
T
ϵ
⃗
=
(
Y
⃗
−
Y
⃗
^
)
T
(
Y
⃗
−
Y
⃗
^
)
=
(
Y
⃗
−
X
⃗
β
)
T
(
Y
⃗
−
X
⃗
β
)
=
Y
⃗
T
Y
⃗
−
2
β
X
⃗
T
Y
⃗
+
β
2
X
⃗
T
X
⃗
\begin{aligned} &\min_{\beta} F(\beta) \\ = &\vec{\epsilon}^T \vec{\epsilon} \\ = &(\vec{Y} - \hat{\vec{Y}})^T (\vec{Y} - \hat{\vec{Y}}) \\ = &(\vec{Y} - \vec{X} \beta)^T (\vec{Y} - \vec{X} \beta) \\ = &\vec{Y}^T \vec{Y} - 2 \beta \vec{X}^T \vec{Y} + \beta^2 \vec{X}^T \vec{X} \end{aligned}
====βminF(β)ϵTϵ(Y−Y^)T(Y−Y^)(Y−Xβ)T(Y−Xβ)YTY−2βXTY+β2XTX
一阶导数
∂
F
∂
β
=
−
2
X
⃗
T
Y
⃗
+
2
β
X
⃗
T
X
⃗
\begin{aligned} \frac{\partial F}{\partial \beta} = -2 \vec{X}^T \vec{Y} + 2 \beta \vec{X}^T \vec{X} \end{aligned}
∂β∂F=−2XTY+2βXTX
二阶导数
∂
2
F
∂
β
2
=
2
X
⃗
T
X
⃗
\begin{aligned} \frac{\partial^2 F}{\partial \beta^2} = 2 \vec{X}^T \vec{X} \end{aligned}
∂β2∂2F=2XTX
当二阶导大于0时,由一阶导等于0得到
β
^
=
(
X
⃗
T
X
⃗
)
−
1
X
⃗
T
Y
⃗
\hat{\beta} = (\vec{X}^T \vec{X})^{-1} \vec{X}^T \vec{Y}
β^=(XTX)−1XTY
真实值
β
\beta
β是常量,估计量
β
^
\hat{\beta}
β^是随机变量。样本量越大,
β
^
\hat{\beta}
β^越接近
β
\beta
β
均值
E
(
β
^
)
=
β
\mathbb{E}(\hat{\beta})=\beta
E(β^)=β
方差(注意其中
X
⃗
\vec{X}
X是已知的,残差项
ϵ
⃗
\vec{\epsilon}
ϵ是随机变量,且均值为0,方差为
s
2
s^2
s2)
E
[
(
β
^
−
β
)
2
]
=
E
[
(
β
−
(
X
⃗
T
X
⃗
)
−
1
X
⃗
T
Y
⃗
)
(
(
β
−
(
X
⃗
T
X
⃗
)
−
1
X
⃗
T
Y
⃗
)
T
]
=
E
[
(
β
−
(
X
⃗
T
X
⃗
)
−
1
X
⃗
T
(
X
⃗
β
+
ϵ
⃗
)
)
(
(
β
−
(
X
⃗
T
X
⃗
)
−
1
X
⃗
T
(
X
⃗
β
+
ϵ
⃗
)
)
T
]
=
E
[
(
X
⃗
T
X
⃗
)
−
1
X
T
ϵ
⃗
ϵ
⃗
T
X
(
X
⃗
T
X
⃗
)
−
1
]
=
E
[
(
X
⃗
T
X
⃗
)
−
1
X
T
ϵ
⃗
ϵ
⃗
T
X
(
X
⃗
T
X
⃗
)
−
1
]
=
(
X
⃗
T
X
⃗
)
−
1
X
T
E
[
ϵ
⃗
ϵ
⃗
T
]
X
(
X
⃗
T
X
⃗
)
−
1
=
(
X
⃗
T
X
⃗
)
−
1
X
T
s
2
I
⃗
X
(
X
⃗
T
X
⃗
)
−
1
=
s
2
(
X
⃗
T
X
⃗
)
−
1
\begin{aligned} &\mathbb{E}[(\hat{\beta} - \beta)^2] \\ = &\mathbb{E}[(\beta - (\vec{X}^T \vec{X})^{-1} \vec{X}^T \vec{Y})((\beta - (\vec{X}^T \vec{X})^{-1} \vec{X}^T \vec{Y})^T] \\ = &\mathbb{E}[(\beta - (\vec{X}^T \vec{X})^{-1} \vec{X}^T (\vec{X} \beta + \vec{\epsilon}))((\beta - (\vec{X}^T \vec{X})^{-1} \vec{X}^T (\vec{X} \beta + \vec{\epsilon}))^T] \\ = &\mathbb{E}[ (\vec{X}^T \vec{X})^{-1} X^T \vec{\epsilon} \vec{\epsilon}^T X (\vec{X}^T \vec{X})^{-1}] \\ = &\mathbb{E}[ (\vec{X}^T \vec{X})^{-1} X^T \vec{\epsilon} \vec{\epsilon}^T X (\vec{X}^T \vec{X})^{-1}] \\ = &(\vec{X}^T \vec{X})^{-1} X^T \mathbb{E}[\vec{\epsilon} \vec{\epsilon}^T] X (\vec{X}^T \vec{X})^{-1} \\ = &(\vec{X}^T \vec{X})^{-1} X^T s^2 \vec{I} X (\vec{X}^T \vec{X})^{-1} \\ = & s^2 (\vec{X}^T \vec{X})^{-1} \end{aligned}
=======E[(β^−β)2]E[(β−(XTX)−1XTY)((β−(XTX)−1XTY)T]E[(β−(XTX)−1XT(Xβ+ϵ))((β−(XTX)−1XT(Xβ+ϵ))T]E[(XTX)−1XTϵϵTX(XTX)−1]E[(XTX)−1XTϵϵTX(XTX)−1](XTX)−1XTE[ϵϵT]X(XTX)−1(XTX)−1XTs2IX(XTX)−1s2(XTX)−1
多因子模型
R A = α + ∑ j = 1 m β j × F j + ϵ R^A = \alpha + \sum_{j=1}^{m} \beta^j \times F^j + \epsilon RA=α+j=1∑mβj×Fj+ϵ
- α \alpha α是截距项
- F 1 , ⋯ , F m F^1, \cdots, F^m F1,⋯,Fm是m个因子,对应的斜率系数为 β 1 , ⋯ , β m \beta^1, \cdots, \beta^m β1,⋯,βm
总共需要估计 m + 1 m+1 m+1个参数
模型仍然是 Y ⃗ = X ⃗ β + ϵ ⃗ \vec{Y} = \vec{X} \beta + \vec{\epsilon} Y=Xβ+ϵ
因变量
Y
⃗
=
(
Y
1
⋮
Y
i
⋮
Y
n
)
=
(
R
1
A
⋮
R
i
A
⋮
R
n
A
)
\vec{Y} = \left( \begin{matrix} Y _1 \\ \vdots \\ Y _i \\ \vdots \\ Y _n \\ \end{matrix} \right) = \left( \begin{matrix} R_1^A \\ \vdots \\ R_i^A \\ \vdots \\ R_n^A \\ \end{matrix} \right)
Y=⎝⎜⎜⎜⎜⎜⎜⎛Y1⋮Yi⋮Yn⎠⎟⎟⎟⎟⎟⎟⎞=⎝⎜⎜⎜⎜⎜⎜⎛R1A⋮RiA⋮RnA⎠⎟⎟⎟⎟⎟⎟⎞
自变量
X
⃗
=
(
F
1
1
⋯
F
1
j
⋯
F
1
m
1
⋮
⋮
⋮
⋮
⋮
⋮
F
i
1
⋯
F
i
j
⋯
F
i
m
1
⋮
⋮
⋮
⋮
⋮
⋮
F
n
1
⋯
F
n
j
⋯
F
n
m
1
)
\vec{X} = \left( \begin{matrix} F_1^1 & \cdots & F_1^j & \cdots & F_1^m & 1 \\ \vdots & \vdots & \vdots & \vdots & \vdots & \vdots \\ F_i^1 & \cdots & F_i^j & \cdots & F_i^m & 1 \\ \vdots & \vdots & \vdots & \vdots & \vdots & \vdots \\ F_n^1 & \cdots & F_n^j & \cdots & F_n^m & 1 \\ \end{matrix} \right)
X=⎝⎜⎜⎜⎜⎜⎜⎛F11⋮Fi1⋮Fn1⋯⋮⋯⋮⋯F1j⋮Fij⋮Fnj⋯⋮⋯⋮⋯F1m⋮Fim⋮Fnm1⋮1⋮1⎠⎟⎟⎟⎟⎟⎟⎞
系数
β
⃗
=
(
β
1
⋮
β
i
⋮
β
m
α
)
\vec{\beta} = \left( \begin{matrix} \beta^1 \\ \vdots \\ \beta^i \\ \vdots \\ \beta^m \\ \alpha \\ \end{matrix} \right)
β=⎝⎜⎜⎜⎜⎜⎜⎜⎜⎛β1⋮βi⋮βmα⎠⎟⎟⎟⎟⎟⎟⎟⎟⎞
估计值仍然是 β ⃗ ^ = ( X ⃗ T X ⃗ ) − 1 X ⃗ T Y ⃗ \hat{\vec{\beta}} = (\vec{X}^T \vec{X})^{-1} \vec{X}^T \vec{Y} β^=(XTX)−1XTY
协方差矩阵 E [ ( β ⃗ ^ − β ⃗ ) ( β ⃗ ^ − β ⃗ ) T ] = s 2 ( X ⃗ T X ⃗ ) − 1 \mathbb{E}[(\hat{\vec{\beta}} - \vec{\beta})(\hat{\vec{\beta}} - \vec{\beta})^T] = s^2 (\vec{X}^T \vec{X})^{-1} E[(β^−β)(β^−β)T]=s2(XTX)−1
2.2.3 等式约束条件的优化问题
等式约束条件的优化问题的一般形式:
min
x
1
,
⋯
,
x
n
f
(
x
1
,
⋯
,
x
n
)
\begin{aligned} \min_{x_1, \cdots, x_n} f(x_1,\cdots, x_n) \end{aligned}
x1,⋯,xnminf(x1,⋯,xn)
subject to :
g
1
(
x
1
,
⋯
,
x
n
)
=
b
1
⋮
g
m
(
x
1
,
⋯
,
x
n
)
=
b
m
\begin{aligned} & g_1(x_1,\cdots, x_n) & = & \ b_1 \\ & & \vdots \\ & g_m(x_1,\cdots, x_n) & = & \ b_m \\ \end{aligned}
g1(x1,⋯,xn)gm(x1,⋯,xn)=⋮= b1 bm
这类问题无法使用标准微积分方法求解。
拉格朗日方法和拉格朗日乘子
L
(
x
⃗
,
λ
⃗
)
=
f
(
x
⃗
)
+
∑
j
=
1
m
λ
j
(
g
j
(
x
⃗
)
−
b
j
)
L(\vec{x}, \vec{\lambda}) = f(\vec{x}) + \sum_{j=1}^{m} \lambda_j (g_j(\vec{x})-b_j)
L(x,λ)=f(x)+j=1∑mλj(gj(x)−bj)
符号 x ⃗ = ( x 1 , ⋯ , x n ) , λ ⃗ = ( λ 1 , ⋯ , λ m ) \vec{x}=(x_1,\cdots,x_n), \ \vec{\lambda}=(\lambda_1, \cdots, \lambda_m) x=(x1,⋯,xn), λ=(λ1,⋯,λm)
拉格朗日方程将n变量m个等式约束条件的问题变换为n+m变量无约束条件的优化问题
min
x
⃗
,
λ
⃗
L
(
x
⃗
,
λ
⃗
)
\min_{\vec{x}, \vec{\lambda}} L(\vec{x}, \vec{\lambda})
x,λminL(x,λ)
一阶条件
∂
L
∂
x
i
(
x
)
=
∂
f
∂
x
i
(
x
)
+
∑
j
=
1
m
λ
j
∂
g
j
∂
x
i
(
x
)
=
0
,
i
=
1
,
⋯
,
n
∂
L
∂
λ
j
(
x
)
=
g
j
(
x
)
−
b
j
=
0
,
j
=
1
,
⋯
,
m
\begin{aligned} & \frac{\partial L}{\partial x_i}(x) = \frac{\partial f}{\partial x_i}(x) + \sum_{j=1}^{m} \lambda_j \frac {\partial g_j}{\partial x_i}(x) = 0, \ i = 1, \cdots, n \\ & \frac{\partial L}{\partial \lambda_j}(x) = g_j(x) - b_j = 0, \ j = 1, \cdots, m \end{aligned}
∂xi∂L(x)=∂xi∂f(x)+j=1∑mλj∂xi∂gj(x)=0, i=1,⋯,n∂λj∂L(x)=gj(x)−bj=0, j=1,⋯,m
通过一阶条件求解之后,还需要检查海森条件(二阶条件)
最小化包含N的风险资产的资产组合的风险
要求解的问题:在给定回报的情况下,最小化组合风险
目标函数
min
w
⃗
=
1
2
σ
π
2
=
1
2
w
⃗
T
Σ
⃗
w
⃗
\begin{aligned} \min_{\vec{w}} = \frac{1}{2} \sigma_{\pi}^2 = \frac{1}{2} \vec{w}^T \vec{\Sigma} \vec{w} \end{aligned}
wmin=21σπ2=21wTΣw
约束条件
- μ π = μ ⃗ T w ⃗ = w ⃗ T μ ⃗ = m \mu_{\pi} = \vec{\mu}^T \vec{w} = \vec{w}^T \vec{\mu} = m μπ=μTw=wTμ=m
- 1 ⃗ T w ⃗ = w ⃗ T 1 ⃗ = 1 \vec{1}^T \vec{w} = \vec{w}^T \vec{1} = 1 1Tw=wT1=1
使用拉格朗日方法,引入拉格朗日乘子
L
(
w
⃗
,
λ
,
γ
)
=
1
2
w
⃗
T
Σ
⃗
w
⃗
+
λ
(
m
−
w
⃗
T
μ
)
+
γ
(
1
−
w
⃗
T
1
⃗
)
L(\vec{w}, \lambda, \gamma) = \frac{1}{2} \vec{w}^T \vec{\Sigma} \vec{w} + \lambda(m - \vec{w}^T \mu) + \gamma(1 - \vec{w}^T \vec{1})
L(w,λ,γ)=21wTΣw+λ(m−wTμ)+γ(1−wT1)
一阶条件
∂
L
∂
w
(
w
,
λ
,
γ
)
=
Σ
⃗
w
⃗
−
λ
μ
⃗
−
γ
1
⃗
\frac{\partial L}{\partial w} (w, \lambda, \gamma) = \vec{\Sigma} \vec{w} - \lambda \vec{\mu} - \gamma \vec{1}
∂w∂L(w,λ,γ)=Σw−λμ−γ1
求解得到
w
⃗
∗
=
Σ
⃗
−
1
(
λ
μ
⃗
+
γ
1
⃗
)
\vec{w}^* = \vec{\Sigma}^{-1} (\lambda \vec{\mu} + \gamma \vec{1})
w∗=Σ−1(λμ+γ1)
带入到约束条件
m
=
μ
⃗
T
w
⃗
=
μ
⃗
T
Σ
⃗
−
1
(
λ
μ
⃗
+
γ
1
⃗
)
=
λ
μ
⃗
T
Σ
⃗
−
1
μ
⃗
+
γ
μ
⃗
T
Σ
⃗
−
1
1
⃗
1
=
1
⃗
T
w
⃗
=
1
⃗
T
Σ
⃗
−
1
(
λ
μ
⃗
+
γ
1
⃗
)
=
λ
1
⃗
T
Σ
⃗
−
1
μ
⃗
+
γ
1
⃗
T
Σ
⃗
−
1
1
⃗
m = \vec{\mu}^T \vec{w} = \vec{\mu}^T \vec{\Sigma}^{-1} (\lambda \vec{\mu} + \gamma \vec{1}) = \lambda \vec{\mu}^T \vec{\Sigma}^{-1} \vec{\mu} + \gamma \vec{\mu}^T \vec{\Sigma}^{-1} \vec{1} \\ 1 = \vec{1}^T \vec{w} = \vec{1}^T \vec{\Sigma}^{-1} (\lambda \vec{\mu} + \gamma \vec{1}) = \lambda \vec{1}^T \vec{\Sigma}^{-1} \vec{\mu} + \gamma \vec{1}^T \vec{\Sigma}^{-1} \vec{1}
m=μTw=μTΣ−1(λμ+γ1)=λμTΣ−1μ+γμTΣ−111=1Tw=1TΣ−1(λμ+γ1)=λ1TΣ−1μ+γ1TΣ−11
定义以下标量
{
A
=
1
⃗
T
Σ
⃗
−
1
1
⃗
B
=
1
⃗
T
Σ
⃗
−
1
μ
⃗
=
μ
⃗
T
Σ
⃗
−
1
1
⃗
C
=
μ
⃗
T
Σ
⃗
−
1
μ
⃗
\left\{\begin{aligned} A &= \vec{1}^T \vec{\Sigma}^{-1} \vec{1} \\ B &= \vec{1}^T \vec{\Sigma}^{-1} \vec{\mu} = \vec{\mu}^T \vec{\Sigma}^{-1} \vec{1} \\ C &= \vec{\mu}^T \vec{\Sigma}^{-1} \vec{\mu} \\ \end{aligned}\right.
⎩⎪⎪⎨⎪⎪⎧ABC=1TΣ−11=1TΣ−1μ=μTΣ−11=μTΣ−1μ
存在
A
C
−
B
2
>
0
AC-B^2 \gt 0
AC−B2>0
C λ + B γ = m B λ + A γ = 1 \left.\begin{aligned} C \lambda + B \gamma = m \\ B \lambda + A \gamma = 1 \\ \end{aligned}\right. Cλ+Bγ=mBλ+Aγ=1
求解得到
{
λ
=
A
m
−
B
A
C
−
B
2
γ
=
C
−
B
m
A
C
−
B
2
\left\{\begin{aligned} \lambda &= \frac{Am-B}{AC-B^2} \\ \gamma &= \frac{C-Bm}{AC-B^2} \\ \end{aligned}\right.
⎩⎪⎪⎨⎪⎪⎧λγ=AC−B2Am−B=AC−B2C−Bm
代入可得到
w
⃗
∗
\vec{w}^*
w∗的解析式
最小方差组合
将权重带入后,将方差表示为 m m m的函数 σ π 2 ( m ) = A m 2 − 2 B m + C A C − B 2 \sigma_{\pi}^2(m) = \frac{Am^2 - 2Bm + C}{AC-B^2} σπ2(m)=AC−B2Am2−2Bm+C
求解最小方差组合就是求解 min m σ π 2 ( m ) \min_m \sigma_{\pi}^2 (m) mminσπ2(m)
对m求导得到 m g = B A m_g = \frac{B}{A} mg=AB
代入得到
权重
w
g
=
Σ
⃗
−
1
1
⃗
A
\begin{aligned} w_g = \frac{\vec{\Sigma}^{-1} \vec{1}}{A} \end{aligned}
wg=AΣ−11
方差 σ g 2 = 1 A \begin{aligned} \sigma_g^2 = \frac{1}{A} \end{aligned} σg2=A1
N个风险资产和无风险资产
优化问题变为:
目标函数
min
w
⃗
=
1
2
σ
π
2
=
1
2
w
⃗
T
Σ
⃗
w
⃗
\begin{aligned} \min_{\vec{w}} = \frac{1}{2} \sigma_{\pi}^2 = \frac{1}{2} \vec{w}^T \vec{\Sigma} \vec{w} \end{aligned}
wmin=21σπ2=21wTΣw
约束条件
- r + w ⃗ T ( μ ⃗ − r 1 ⃗ ) = m r + \vec{w}^T (\vec{\mu} - r \vec{1})= m r+wT(μ−r1)=m
拉格朗日方程 V ( ω ⃗ , λ ) = 1 2 ω ⃗ T Σ ⃗ ω ⃗ + λ ( r + ω ⃗ T ( μ ⃗ − r 1 ⃗ ) − m ) \begin{aligned} V(\vec{\omega}, \lambda) = \frac{1}{2} \vec{\omega}^T \vec{\Sigma} \vec{\omega} + \lambda (r + \vec{\omega}^T (\vec{\mu} - r \vec{1}) - m) \end{aligned} V(ω,λ)=21ωTΣω+λ(r+ωT(μ−r1)−m)
一阶导(梯度)
∂ V ∂ ω ⃗ ( ω ⃗ ∗ , λ ) = Σ ⃗ ω ⃗ ∗ + λ ( μ ⃗ − r 1 ⃗ ) = 0 → ω ⃗ ∗ = λ Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) \begin{aligned} & \frac {\partial V}{\partial \vec{\omega}}(\vec{\omega} ^ *, \lambda) = \vec{\Sigma} \vec{\omega}^* + \lambda (\vec{\mu} - r \vec{1}) = 0 \\ \to \quad & \vec{\omega}^* = \lambda \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1}) \\ \end{aligned} →∂ω∂V(ω∗,λ)=Σω∗+λ(μ−r1)=0ω∗=λΣ−1(μ−r1)
将 ω ⃗ ∗ \vec{\omega}^* ω∗代入到约束条件
r + [ λ Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) ] T ( μ ⃗ − r 1 ⃗ ) = m → λ ( μ ⃗ − r 1 ⃗ ) T Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) = m − r → λ = m − r ( μ ⃗ − r 1 ⃗ ) T Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) → ω ⃗ ∗ = λ Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) = ( m − r ) Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) ( μ ⃗ − r 1 ⃗ ) T Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) \begin{aligned} & r + [\lambda \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1})]^T (\vec{\mu} - r \vec{1}) = m \\ \to \quad & \lambda (\vec{\mu} - r \vec{1})^T \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1}) = m - r \\ \to \quad & \lambda = \frac{m - r}{(\vec{\mu} - r \vec{1})^T \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1})} \\ \to \quad & \vec{\omega}^* = \lambda \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1}) = \frac {(m -r) \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1})} {(\vec{\mu} - r \vec{1})^T \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1})} \\ \end{aligned} →→→r+[λΣ−1(μ−r1)]T(μ−r1)=mλ(μ−r1)TΣ−1(μ−r1)=m−rλ=(μ−r1)TΣ−1(μ−r1)m−rω∗=λΣ−1(μ−r1)=(μ−r1)TΣ−1(μ−r1)(m−r)Σ−1(μ−r1)
夏普比率
D = ( μ ⃗ − r 1 ⃗ ) T Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) = μ ⃗ T Σ ⃗ − 1 μ ⃗ − r 1 ⃗ T Σ ⃗ − 1 μ ⃗ − r μ ⃗ T Σ ⃗ − 1 1 ⃗ + r 2 1 ⃗ T Σ ⃗ − 1 1 ⃗ = C − 2 r B + r 2 A 2 ω ⃗ ∗ = ( μ π − r ) Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) D σ π = ω ⃗ ∗ T Σ ⃗ ω ⃗ ∗ = ( μ π − r ) 2 D 2 [ Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) ] T Σ ⃗ [ Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) ] = ( μ π − r ) 2 D 2 ( μ ⃗ − r 1 ⃗ ) T Σ ⃗ − 1 Σ ⃗ Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) = ( μ π − r ) 2 D 2 ( μ ⃗ − r 1 ⃗ ) T Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) = ( μ π − r ) 2 D S R = μ π − r σ π = D = ( μ ⃗ − r 1 ⃗ ) T Σ ⃗ − 1 ( μ ⃗ − r 1 ⃗ ) \begin{aligned} D =& (\vec{\mu} - r \vec{1})^T \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1}) \\ = & \vec{\mu}^T \vec{\Sigma}^{-1} \vec{\mu} - r \vec{1}^T \vec{\Sigma}^{-1} \vec{\mu} - r \vec{\mu}^T \vec{\Sigma}^{-1} \vec{1} + r^2 \vec{1}^T \vec{\Sigma}^{-1} \vec{1} \\ = &C - 2rB + r^2 A^2 \\ \\ \vec{\omega}^* =& \frac {(\mu_{\pi} - r) \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1})} {D} \\ \\ \sigma_{\pi} =& \vec{\omega}^{*T} \vec{\Sigma} \vec{\omega}^* \\ =& \frac{(\mu_{\pi} - r)^2}{D^2} [\vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1})]^T \vec{\Sigma} [\vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1}) ] \\ =& \frac{(\mu_{\pi} - r)^2}{D^2} (\vec{\mu} - r \vec{1})^T \vec{\Sigma}^{-1} \vec{\Sigma} \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1}) \\ =& \frac{(\mu_{\pi} - r)^2}{D^2} (\vec{\mu} - r \vec{1})^T \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1}) \\ =& \frac{(\mu_{\pi} - r)^2}{D} \\ \\ SR =& \frac {\mu_{\pi} - r}{\sigma_{\pi}} \\ =& \sqrt{D} \\ =& \sqrt{(\vec{\mu} - r \vec{1})^T \vec{\Sigma}^{-1} (\vec{\mu} - r \vec{1})} \end{aligned} D===ω∗=σπ=====SR===(μ−r1)TΣ−1(μ−r1)μTΣ−1μ−r1TΣ−1μ−rμTΣ−11+r21TΣ−11C−2rB+r2A2D(μπ−r)Σ−1(μ−r1)ω∗TΣω∗D2(μπ−r)2[Σ−1(μ−r1)]TΣ[Σ−1(μ−r1)]D2(μπ−r)2(μ−r1)TΣ−1ΣΣ−1(μ−r1)D2(μπ−r)2(μ−r1)TΣ−1(μ−r1)D(μπ−r)2σπμπ−rD(μ−r1)TΣ−1(μ−r1)
- 矩阵运算满足结合律
- Σ ⃗ \vec{\Sigma} Σ是对称阵, Σ ⃗ − 1 \vec{\Sigma}^{-1} Σ−1也是对称阵
加入无风险资产后,投资组合的有限前沿变为直线,斜率为夏普比率SR
切线组合
切线组合全部投资在风险资产上,即
w
t
⃗
T
1
⃗
=
1
\vec{w_t}^T \vec{1} = 1
wtT1=1
w
t
⃗
=
Σ
⃗
−
1
(
μ
⃗
−
r
1
⃗
)
B
−
A
r
\vec{w_t} = \frac{\vec{\Sigma}^{-1}(\vec{\mu} - r \vec{1})}{B-Ar}
wt=B−ArΣ−1(μ−r1)
均值 m t = w t ⃗ T μ ⃗ = C − B r B − A r m_t = \vec{w_t}^T \vec{\mu} = \frac{C-Br}{B-Ar} mt=wtTμ=B−ArC−Br
方差 σ t = w t ⃗ T Σ ⃗ w t ⃗ = C − 2 r B + r 2 A ( B − A r ) 2 \sigma_t = \vec{w_t}^T \vec{\Sigma} \vec{w_t} = \frac{C-2rB +r^2 A}{(B-Ar)^2} σt=wtTΣwt=(B−Ar)2C−2rB+r2A
2.2.4 Black-Litterman模型
动态资产选择模型,关注超额收益
超额收益 R ~ = R − r 1 ⃗ , R ~ ∼ N ( μ ~ , Σ ) \tilde{R} = R - r \vec{1}, \quad \tilde{R} \sim N(\tilde{\mu}, \Sigma) R~=R−r1,R~∼N(μ~,Σ)
其中 μ ~ = μ ⃗ − r 1 ⃗ \tilde{\mu} = \vec{\mu} - r \vec{1} μ~=μ−r1, Σ \Sigma Σ是协方差矩阵
Black-Litterman模型认为:真实的超额回报均值
μ
~
\tilde{\mu}
μ~无法直接计算或观测到,因此依赖于估计量
π
\pi
π
π
∼
N
(
μ
~
,
Σ
π
)
\pi \sim N(\tilde{\mu}, \Sigma_{\pi})
π∼N(μ~,Σπ)
π
\pi
π是围绕
μ
~
\tilde{\mu}
μ~波动的噪声项(随机项)
π
=
μ
~
+
ϵ
,
ϵ
∼
N
(
0
,
Σ
π
)
\pi = \tilde{\mu} + \epsilon , \quad \epsilon \sim N(0, \Sigma_{\pi})
π=μ~+ϵ,ϵ∼N(0,Σπ)
假设
ϵ
\epsilon
ϵ与
μ
~
\tilde{\mu}
μ~不相关,则
R
~
∼
N
(
μ
~
,
Σ
r
)
,
Σ
r
=
Σ
+
Σ
π
\tilde{R} \sim N(\tilde{\mu}, \Sigma_r) , \quad \Sigma_r = \Sigma + \Sigma_{\pi}
R~∼N(μ~,Σr),Σr=Σ+Σπ
Black-Litterman公式有两种推导方法:
- Theil’s mixed estimation model:使用generalised least squares regression (GLS)方法,GLS是OLS的扩展
- Bayes’s Formula:用获得的新信息优化资产配置,可以用贝叶斯统计方法解释(贝叶斯公式用新条件更新先验概率,得到后验概率)
贝叶斯方法的内容:
- Step 0: Introducing Bayes’ Formula;
- Step 1: Reverse optimization to get the prior;
- Step 2: Inputing the view;
- Step 3: Combining to get the posterior distribution;
- Step 4: Asset allocation.
Step 0: Introducing Bayes’ Formula
P ( E ∣ I ) = P ( I ∣ E ) P ( I ) × P ( E ) P(E|I) = \frac{P(I|E)}{P(I)} \times P(E) P(E∣I)=P(I)P(I∣E)×P(E)
其中:
- E E E是事件
- I I I是新信息
- P ( E ) P(E) P(E)是先验概率,没有信息 I I I之前的概率
- P ( E ∣ I ) P(E|I) P(E∣I)是后验概率,知道信息 I I I之后的概率(主观概率)
- P ( I ) P(I) P(I)是归一化常量
Step 1: Reverse optimization to get the prior
可能的先验条件:
- 平均分配(每个资产 1 / N 1/N 1/N)
- 全局最小方差组合
- any other `neutral and uninformed’ portfolio
这些选择都不理想:
- 得到全局最小方差,隐含了要估计超额收益。如果使用历史收益,可能不能反映未来期望,导致不切实际的组合
- 平均分配不能建立超额回报的先验视角(用均价分配表示什么都没有知道,没有先验知识)
Black-Litterman使用equilibrium CAPM portfolio也就是求解:
max
w
⃗
w
⃗
T
R
~
−
λ
2
w
⃗
T
Σ
w
⃗
\max_{\vec{w}} \vec{w}^T \tilde{R} - \frac{\lambda}{2} \vec{w}^T \Sigma \vec{w}
wmaxwTR~−2λwTΣw
Σ \Sigma Σ是超额回报的协方差矩阵
求解得到 w ⃗ ∗ = 1 λ Σ − 1 R ~ \vec{w}^{*} = \frac{1}{\lambda} \Sigma^{-1} \tilde{R} w∗=λ1Σ−1R~
现在的问题:不知道超额回报向量
对于市场组合有(这里的意思似乎是用市场组合的超额收益
Π
⃗
\vec{\Pi}
Π代替
R
~
\tilde{R}
R~):
w
⃗
M
=
1
λ
Σ
−
1
Π
⃗
Π
⃗
=
λ
Σ
w
⃗
M
\vec{w}_M = \frac{1}{\lambda} \Sigma^{-1} \vec{\Pi} \\ \vec{\Pi} = \lambda \Sigma \vec{w}_M
wM=λ1Σ−1ΠΠ=λΣwM
风险厌恶因子
w ⃗ T Π ⃗ = λ w ⃗ T Σ w ⃗ M = λ σ M 2 \vec{w}^{T} \vec{\Pi} = \lambda \vec{w}^{T} \Sigma \vec{w}_M = \lambda \sigma_M^2 wTΠ=λwTΣwM=λσM2
λ = w ⃗ T Π ⃗ σ M 2 = 1 σ M S M \lambda = \frac{\vec{w}^{T} \vec{\Pi}}{\sigma_M^2} = \frac{1}{\sigma_M} S_M λ=σM2wTΠ=σM1SM
夏普比率
S
M
=
w
⃗
T
Π
⃗
σ
M
S_M = \frac{\vec{w}^{T} \vec{\Pi}}{\sigma_M}
SM=σMwTΠ
在模型中,夏普比例取0.5
假设 Σ Π = τ Σ \Sigma_{\Pi} = \tau \Sigma ΣΠ=τΣ
得到先验概率分布
P
(
E
)
∼
N
(
Π
,
τ
Σ
)
P(E) \sim N(\Pi, \tau \Sigma)
P(E)∼N(Π,τΣ)
参数 τ \tau τ
通常
τ
\tau
τ接近于0,取值位于0.01到0.05之间
如果估计长度为T的时间序列的历史数据,取
τ
≈
T
−
1
\tau \approx T^{-1}
τ≈T−1
Step 2: Inputing the views
- Relative view:用资产之间的收益差衡量
- Absolute view:用资产的绝对收益衡量
P ( I ∣ E ) ∼ N ( Q , Ω ) P(I|E) \sim N(Q, \Omega) P(I∣E)∼N(Q,Ω)
Q
Q
Q是收益率向量
Ω
\Omega
Ω是置信度协方差矩阵
确定协方差矩阵
Ω
\Omega
Ω
Ω
:
=
D
i
a
g
(
P
(
τ
Σ
)
P
T
)
\Omega := Diag(P(\tau \Sigma) P^T)
Ω:=Diag(P(τΣ)PT)
Diag是对角化函数
Step 3: Combining to get the posterior distribution
μ = [ ( τ Σ ) − 1 + P T Ω − 1 P ] − 1 [ ( τ Σ ) − 1 Π + P T Ω − 1 Q ] σ 2 = [ ( τ Σ ) − 1 + P T Ω − 1 P ] − 1 P ( E ∣ I ) ∼ N ( μ , σ 2 ) R ^ I : = E [ R ~ ∣ I ] = μ \begin{aligned} \mu = &[(\tau \Sigma)^{-1} + P^T \Omega^{-1} P]^{-1}[(\tau \Sigma)^{-1} \Pi + P^T\Omega^{-1} Q] \\ \sigma^2 = &[(\tau \Sigma)^{-1} + P^T \Omega^{-1} P]^{-1} \\ P(E|I) \sim &N(\mu, \sigma^2) \\ \hat{R}_I := &\mathbb{E}[\tilde{R}|I] = \mu \end{aligned} μ=σ2=P(E∣I)∼R^I:=[(τΣ)−1+PTΩ−1P]−1[(τΣ)−1Π+PTΩ−1Q][(τΣ)−1+PTΩ−1P]−1N(μ,σ2)E[R~∣I]=μ
Step 4: Asset allocation
max w ⃗ [ r + w T R ^ I ] − λ w T Σ w \max_{\vec{w}} [r + w^T \hat{R}_I] - \frac{\lambda}w^T \Sigma w wmax[r+wTR^I]−wλTΣw
求解得到
w
∗
=
1
λ
Σ
−
1
R
^
I
w^* = \frac{1}{\lambda} \Sigma^{-1} \hat{R}_I
w∗=λ1Σ−1R^I
2.2.5 不等式约束条件:Kuhn-Tucker条件
不等式约束条件的优化问题的一般形式:
min
x
1
,
⋯
,
x
n
f
(
x
1
,
⋯
,
x
n
)
\begin{aligned} \min_{x_1, \cdots, x_n} f(x_1,\cdots, x_n) \end{aligned}
x1,⋯,xnminf(x1,⋯,xn)
subject to :
g
1
(
x
1
,
⋯
,
x
n
)
≤
b
1
⋮
g
m
(
x
1
,
⋯
,
x
n
)
≤
b
m
\begin{aligned} & g_1(x_1,\cdots, x_n) & \le & \ b_1 \\ & & \vdots \\ & g_m(x_1,\cdots, x_n) & \le & \ b_m \\ \end{aligned}
g1(x1,⋯,xn)gm(x1,⋯,xn)≤⋮≤ b1 bm
Kuhn-Tucker条件
λ 0 ∂ f ( x ) ∂ x i + ∑ j = 1 m λ j ∂ ( g j ( x ) − b j ) ∂ x i = 0 i = 1 , ⋯ , n λ j ( g j ( x ) − b j ) = 0 j = 1 , ⋯ , m λ j ≥ 0 j = 0 , ⋯ , m \begin{aligned} \lambda_0 \frac{\partial f(x)}{\partial x_i} + \sum_{j=1}^{m} \lambda_j \frac{\partial (g_j(x) - b_j)}{\partial x_i} &= 0 \qquad i = 1, \cdots, n \\ \lambda_j (g_j(x) - b_j) &= 0 \qquad j = 1, \cdots, m \\ \lambda_j &\ge 0 \qquad j = 0, \cdots, m \\ \end{aligned} λ0∂xi∂f(x)+j=1∑mλj∂xi∂(gj(x)−bj)λj(gj(x)−bj)λj=0i=1,⋯,n=0j=1,⋯,m≥0j=0,⋯,m
Kuhn-Tucker条件在目标函数是凸函数且约束条件为线性时才是充分必要条件。